While current approaches to ontology mapping produce good results by mainly relying on label and structure based similarity measures, there are several cases in which they fail to discover important mappings. In this paper we describe a novel approach to ontology mapping, which is able to avoid this limitation by using background knowledge. Existing approaches relying on background knowledge typically have one or both of two key limitations: 1) they rely on a manually selected reference ontology; 2) they suffer from the noise introduced by the use of semi-structured sources, such as text corpora. Our technique circumvents these limitations by exploiting the increasing amount of semantic resources available online. As a result, there is no need either for a manually selected reference ontology (the relevant ontologies are dynamically selected from an online ontology repository), or for transforming background knowledge in an ontological form. The promising results from experiments on two real life thesauri indicate both that our approach has a high precision and also that it can find mappings, which are typically missed by existing approaches.
Ontology mapping techniques are essential for building semantic bridges between
ontologies. However, current approaches suffer from a number of problems. First
of all, most approaches do not provide a formal semantics to the mapping
structures they produce1. As a result it is difficult for reasoners to make use of these
structures, e.g., to answer queries across ontologies [
A few approaches [
The recent growth of the Semantic Web has resulted in an increased amount
of online available semantic data and has led to the first search engine to exploit
this data, Swoogle [
In this paper we build on these ideas and we describe an approach to ontology mapping which goes beyond similarity-based algorithms by dynamically locating and using relevant background knowledge (Section 3). Since our method derives mappings between pairs of concepts, it can be used to map semantic structures ranging from shallow thesauri to clearly formalized ontologies. We start by discussing in detail the importance of background knowledge in ontology mapping. 2
We discuss two major limitations of current ontology mapping approaches that only rely on label and structure similarity (i.e., syntactic approaches), namely that they don’t provide semantic mappings (Section 2.1) and that they fail to discover some correct mappings when the mapped ontologies are syntactically dissimilar (Section 2.2). We then point out that while the use of background knowledge can be a solution to these limitations, existing approaches using such knowledge have their shortcomings (Section 2.3). 2.1
Semantic Web tools, such as PowerAqua, which wish to reason on the results
of mapping techniques require that the discovered mappings are expressed as
semantic relations between the entities of the ontologies. Formal mapping
languages, such as C-OWL [
An analysis of the state of the art of mapping systems presented in [
As already observed by [
The previous sections suggest that the meaning of the mapped concepts should
be considered to discover meaningful and syntactically unidentifiable mappings.
Unfortunately, while meaning on the Semantic Web is expressed using ontologies,
in the case of ontology mapping, the constituents of a mapping can only be given
meaning in the context of their own distinct ontology, which cannot cover both
the source and target elements, as well as the relation linking them. In other
words, a semantic mapping between two ontologies could only be interpreted in
a larger domain than the ones of these ontologies. Therefore, in order to achieve
semantic mapping, the integration of external knowledge is required as a way to
cover both input ontologies and to fill the semantic gap between them. So far
the following types of background knowledge have been used in mapping:
1. WordNet is one of the most often used sources of background knowledge. For
example, CTxMatch [
The advantage of these approaches is that they use richly axiomatized
ontologies as background knowledge and therefore guarantee the semantic nature
of the mappings. However, a weakness is that the appropriate reference
ontology needs to be manually selected prior to mapping. As already pointed out,
in many scenarios this approach is unfeasible as we might not know in advance
which terms from which ontologies we may want to map. Even in the cases where
a reference ontology can be manually selected prior to performing the mapping,
there is no guarantee that such an ontology actually exists.
3. Online textual resources can provide an important source of background
knowledge. van Hage et. al [
The strength of this approach is that it reduces the high cost of
establishing adequate background knowledge. Indeed, the background knowledge sources
are dynamically discovered and used [
We conclude that the use of background knowledge overcomes the major limitations of syntactic approaches: it allows obtaining semantic relations even between dissimilar ontologies. However, existing approaches either 1) rely on an a priori selected reference ontology or, if they acquire knowledge dynamically, 2) suffer from the noise introduced by knowledge extraction techniques. As a result, they are not suitable for use by novel Semantic Web tools, such as PowerAqua, which require both that the returned mappings are semantically sound and that the relevant background knowledge is dynamically selected, at run-time. In the next section we describe an approach that fulfills these requirements. 3
Our hypothesis is that the growing amount of online available semantic data which makes up the Semantic Web can be used as a source of background knowledge in ontology mapping in a way that satisfies the requirements identified in the previous section. Indeed, this large-scale, heterogeneous semantic data collection provides formally specified knowledge which is likely to be less faulty than that derived from textual sources and therefore lead to better mappings. Moreover, the size and heterogeneity of the collection makes it possible to dynamically select and combine the appropriate knowledge and to avoid the manual selection of a single, large ontology. In the following we investigate increasingly sophisticated approaches to discover and exploit online available ontologies for mapping. We also provide experimental evidence that such mappings can be obtained.
Experimental Data. We have used the dataset described in [
Implementation Details. We explore our idea by implementing different
mapping strategies on top of the Swoogle’05 ontology search engine [
Notations. Each strategy takes two candidate concept names (A and B) as
an input and returns the discovered mapping between them. The corresponding
concepts in the selected ontology are A’ and B’ (“anchor terms”). We rely on
the description logic syntax for semantic relations occurring between concepts
in an ontology, e.g., A’ v B’ means that A’ is a sub-concept of B’ in a selected
ontology and A’ ⊥ B’ means that A’ and B’ are disjoint. The returned mappings
are expressed using C-OWL [
Our simplest strategy consists in using Swoogle to find ontologies containing concepts with the same names as the candidate concepts and to derive mappings from their relationship in the selected ontologies. Figure 1(b) illustrates this strategy with an example where three ontologies are discovered containing the concepts A’ and B’ with the same names as A and B. The first ontology contains no relation between the anchor concepts, while the other two ontologies contain a subsumption relation. The concrete steps of this strategy are: – if A’ v B’ then derive A −v→ B; – if A’ w B’ then derive A −w→ B; 1. Select ontologies containing concepts A’ and B’ corresponding to A and B; 2. For each resulting ontology:
– if A’ ≡ B’ then derive A −≡→ B;
– if A’ ⊥ B’ then derive A −⊥→ B; 3. If no ontology is found, no mapping is derived;
Even if this strategy seams simple, it leads to several implementation choices, depending on the relative importance given to time performance and accuracy of the mapping mechanism:
rithm would stop as soon as a mapping is discovered. This is the easiest way to deal with the multiple returned ontologies but it assumes that the first discovered relation can be trusted and there is no need to inspect the other ontologies. Note that the first ontology returned by Swoogle does not necessarily contain a relation between the candidate concepts (like in the example Figure 1(b)). Here we use the first ontology containing such a relation, but, in another implementation, it could be considered that if an ontology covers the candidate concepts without relating them, then no mapping should be derived.
Dealing with contradictions. Instead of relying on the information provided by only one ontology as before, we can envisage to combine the results obtained using all the selected ontologies. Mappings resulting from different sources v w can be different (e.g., A −→ B and A −→ B), or, in the worst case, inconsistent v ⊥ (e.g., A −→ B and A −→ B). Several ways of dealing with these contradictions can be considered: we can keep all the mappings (favoring recall), only keep mappings without contradiction (favoring precision), keep the mappings that are derived from most of the ontologies, or try to combine the results (e.g., by deriving A −≡→ B from A −v→ B and A −→ B). In any case, combining the results w from several ontologies is more time consuming (but more reliable) than deriving it from a single ontology.
Considering a particular level of inferences. In the simplest implementation, we can rely on direct and declared relations between A’ and B’ in the selected ontology. But, for better results, indirect and inferred relations should also be exploited (e.g., if A’ v C and C ⊥ B’, then A’ ⊥ B’). Different levels of inferences can be considered (no inference, basic transitivity, DL reasoning), each of them representing a particular compromise between the performance of mapping and the completeness of the result.
Experimental results. For our experiments, we implemented this first strategy using basic transitivity reasoning (i.e., taking into account all parents of A’ and B’) and stopping as soon as a relation was found.
A-Food vs. SR-Meat: We obtained three mappings for these term sets: Beef, P ork, P oultry −v→ F ood. All mappings were derived from the Tap ontology4, where, for example, Beef v ReadM eat v M eatOrP oultry v F ood. A-Animal vs. SR-Meat: For these hierarchies, our implementation yielded in v a single mapping, Bacon −→ P ork, which can be found as is in Tap.
Analyzing our results, we discovered that a key factor in the efficiency of our approach is the level to which the candidate terms are covered by Swoogle. 4 http://139.91.183.30:9090/RDF/VRP/Examples/tap.rdf Indeed, comparing our results to the gold standard mappings, we observed that 24 out of 32 for A-Food vs. SR-Meat (and 20 out of 31 for A-Animal vs. SRMeat) involve concepts that do not exist in any ontology known to Swoogle (e.g., GuineaHen, Quail, Squab). Our experiments are quite strict with respect to finding anchor terms for the candidate concepts: only concepts with identical names are considered. In the next section we suggest ways to reduce this problem. 3.2
In order to discover more ontologies that cover the candidate concepts, the process of finding anchor terms must be more flexible. This flexibility can be achieved by considering the following techniques:
A. String normalization. Differences between concept names can be based on simple differences in naming conventions (e.g., T U RKEY BREAST and T urkeyBreast). Most mapping mechanisms use string normalization techniques, that consist in transforming strings into a standard form before comparison. Our ontology selection relies on such mechanisms as well.
B. Dealing with compound names. Compound names are particularly difficult to match as they are likely to appear under slightly different forms. Several mapping techniques suggest to be more flexible when searching for compound terms and to allow for:
does not appear in Swoogle, but RoastT urkey does.
Additional constituents. For example, T urkeyBreast is not covered but
T urkeyM eatBreast (which additionally contains M eat) is.
Less constituents. Some compound terms are only partially covered. For
example, M eatP roduct does not exist in Swoogle, but M eat does.
Such a flexible matching is also used when discovering anchor terms in the work
of Aleksovski et al. [
such as synonymy can be used to replace terms with their semantic equivalents. A good source for synonymy information is WordNet. However, the drawbacks of WordNet are that it is difficult to get relevant synonyms unless the sense of the term is known a priory and that compound terms are weakly covered.
Experimental results. Just to prove the point that extended coverage can have a significant effect on the obtained mappings, we rerun our experiments by replacing some terms with their syntactic approximates. We replaced T urkeyRoast with RoastT urkey (SR-Meat), T urkeyBreast with T urkeyM eatBreast (SR-Meat) and, M eatP roduct with M eat (A-Animal). v A-Food vs. SR-Meat: We obtained that RoastT urkey −→ P reparedF ood v and RoastT urkey −→ F ood because RoastT urkey v T urkeyDish v P oultryDish v M eatDish v P reparedF ood v F ood. Also, T urkeyM eatBreast −v→ F ood. All mappings were derived from Tap.
A-Animal vs. SR-Meat: We obtained three extra mappings: Beef −v→ M eat5, v v Ham −→ M eat6 and P ork −→ M eat7.
Increasing the coverage of the mapped terms by replacing them with semantically similar variants leads to more mappings. However, another problem comes from the fact that both candidate concepts might not appear in a single ontology, even if each of them appears by itself in many ontologies. Therefore, another way to obtain more mappings is to extend the matching strategy to combine information derived from multiple ontologies, as detailed next. 3.3
The previous strategies (S1 and S2) assume that a semantic relation between the candidate concepts can be discovered in a single ontology. However, some relations could be distributed over several ontologies. Therefore, if no ontology is found that relates both candidate concepts, then the mappings should be derived from two (or more) ontologies. In this strategy, mapping is a recursive task where two concepts can be mapped because the concepts they relate in some ontologies are themselves mapped (Figure 1(c)): 1. If no ontologies are found that contain both A and B then select all ontologies containing a concept A’ corresponding to A; 2. For each of the resulting ontologies: (a) for each C such that A’ v C, search for mappings between C and B; (b) for each C such that A’ w C, search for mappings between C and B; (c) derive mappings using the following rules: – (r1) if A’ v C and C −v→ B then A −v→ B – (r2) if A’ v C and C −≡→ B then A −v→ B – (r3) if A’ v C and C −⊥→ B then A −⊥→ B – (r4) if A’ w C and C −w→ B then A −w→ B – (r5) if A’ w C and C −≡→ B then A −w→ B
In this strategy, steps (a) and (b) can be ran in parallel and stopped when one of them is able to establish a mapping. These two steps correspond to the recursive part of the algorithm. The task of searching for mappings between C and B can be realized using one of our three strategies.
Experimental results. We have implemented this algorithm by using the first mapping strategy (S1) in the recursive part.
A-Food vs. SR-Meat: By combining information available in different ontologies, we obtained that Chicken, Duck, Goose, T urkey −v→ F ood because they are subclasses of P oultry in some ontologies 8 and P oultry v F ood in
Tap (r1). We also discovered that Ham −v→ F ood because Ham v M eat and M eat v F ood in SUMO9 (r1). Finally, we found that Ham −⊥→ Seef ood because Ham v M eat and M eat ⊥ Seaf ood10 (r3).
A-Animal vs. SR-Meat: Because Beef, Ham, P ork v M eat and ⊥ M eat ⊥ Seaf ood we derive that Beef, Ham, P ork −→ Seaf ood (r3). Note that incompatibility mappings are not specified in the gold standard. 4
The aim of this paper was to show the feasibility and the potential advantages of using automatically selected online ontologies as background knowledge for semantic mapping. As Table 1 shows, our experiments on two real life examples have provided promising results, which are consistent with our idea of semantic mappings, as discussed in Section 2. In particular, the output of our algorithm provides mappings, which i) are expressed in terms of semantic relations (subsumption, disjunction); ii) rely on semantics, as expressed in external ontologies; and iii) in many cases would have not been discovered by syntactic techniques (e.g., because the strings denoting similar concepts are very different).
Mappings
T urkeyM eatBreast −v→ F ood S1 S2
A-Food vs. SR-16 Beef, P ork, P oultry −v→ F ood
A-Animal vs. SR-16
Bacon −v→ P ork RoastT urkey −v→ F ood, P reparedF ood Beef, Ham, P ork −v→ M eat
S3 Chicken, Goose, T urkey, Duck −v→ F ood Beef, Ham, P ork −⊥→ Seaf ood* Total
Ham −v→ F ood; Ham −⊥→ Seef ood* 11 (+1*) 4 (+3*)
Note that the technique presented here is not meant to be used in isolation, as
an alternative to current approaches. On the contrary, we plan to integrate our
technique with “syntactic” techniques, to develop a robust and comprehensive
ontology mapping method. For this reason it is difficult to provide a detailed
comparison with other approaches, using standard measures of precision and
recall. Because our technique is meant to enhance, rather than replace existing
methods, it scores very highly on precision but relatively low on recall, about
30% on the test case provided by [
The broader context of our work is one of providing meaningful mappings
that can be used by the next generation of Semantic Web applications [
Another goal of this paper was to identify some of the research issues brought
up by the innovative aspects of our technique. The first innovative aspect is that
appropriate background knowledge is automatically selected from the variety of
ontologies available on the Semantic Web. As a result, important issues for us
concern i) the current level of semantic coverage of the Semantic Web and ii) the
quality of the tools that give access to it. Regarding the second point, although
Swoogle [
The second innovative aspect of our technique, that of combining facts from different ontologies, leads to another important issue: how to deal with contradictions. Online ontologies are made for different purposes, in different contexts and therefore, can lead to contradictory (or inconsistent) mappings. As already mentioned, one of the advantages of semantic techniques is that resulting mappings can be justified and explained. In that sense, one way to deal with contradictions would be to relate mappings to the ontologies on which they are based. Using this solution, contradictory mappings would still co-exist but in a contextualized form, i.e., justified and valid only in the context of some particular ontologies. Another way to deal with contradictions would be to rely only on ontologies sharing a similar context with the mapped ones. Indeed, when trying to map the T urkey concept, we would more likely find relevant mappings using ontologies also containing concepts like F ood or M eat, than ontologies covering countries. This implies that more advanced ontology selection techniques are needed which can consider similarity between ontologies as a selection criterion.
In addition to tackling the aforementioned issues, the experiments presented here need to be followed by several studies. A key next step concerns the complete implementation and evaluation of our technique, which is currently restricted to subsumption and disjunction relations between concepts. Our plan is to extend the technique to also map properties and individuals, so to increase the range of discovered mappings. Finally, when presenting our strategies we emphasized the trade-off between the performance and the accuracy of the mapping mechanism. Finding a good compromise between these two aspects is a hard task and we plan to address this issue by reformulating our technique as an anytime algorithm.
We are grateful to Willem van Hage for providing us with his experimental data set. This work was funded by the AKT project sponsored under grant number GR/N15764/01, and the Open Knowledge and NeOn projects sponsored under EC grant numbers IST-FF6-027253 and IST-FF6-027595.