Semantic heterogeneity caused by the use of di erent ontologies to describe the same topics represents an obstacle for many data integration tasks on the Web of Data, in particular, discovering relevant repositories for interlinking and comparing repositories with respect to the coverage of speci c domains. To facilitate these tasks, mappings between schema terms are needed alongside the links between instances. Currently, explicitly speci ed schema-level mappings are scarce in comparison with instance-level links. However, by analysing existing instance-level links it is possible to capture correspondences between classes to which these instances belong. In our experiments, we applied this approach on a large scale to generate schema-level mappings between several Linked Data repositories. The results of these experiments provide some interesting insights about the use of ontologies on the Web of Data and schema-level relations which emerge from existing data-level interlinks.
One of the main motivations behind large-scale data publishing using the Linked
Data approach [
These questions can partially be answered with the help of meta-level descriptions using the voiD ontology1. However, voiD descriptors may be insu cient to compare some of the characteristics (e.g., whether the domain of food and diets is better covered in Freebase or DBPedia). Moreover, voiD descriptors not always describe all relevant properties of datasets (e.g., dcterms:subject is not always provided) and for some datasets may be not available.
One of the major obstacles which complicate this kind of analysis is schema heterogeneity. It can be di cult to establish automatically that two repositories describe the same kind of data, retrieve relevant data subsets from them, and make a comparison, if these repositories use di erent terminology to describe the same or semantically similar types of instances. For example, a hypothetical repository describing a TV program may need to refer to descriptions of movies, music pieces, and their performers. There are several repositories available on the Web: e.g., speci c sources describing the music topic (MusicBrainz, Jamendo, etc.), the movie topic (LinkedMDB), as well as generic sources covering both (DBPedia, Freebase). In order to compare how well these repositories are suitable as reference sources, it is useful to know which classes in the respective ontologies contain overlapping data: e.g., music:MusicArtist and dbpedia:MusicalArtist, linkedmdb: lm and dbpedia:Movie, etc. Having a high-level overview of schema-level correspondences, which would show the coverage of topics by available ontologies would help the data publisher to make appropriate choices.
In this paper, we described our work on constructing such a network of class level mappings for a subset of the Linked Data cloud. So far, several ontologies used by popular Linked Data repositories were enriched with mappings connecting them to other ontologies (most notably, in the context of the UMBEL project2). However, these mappings, constructed in a top-down way, only cover a limited subset of the Web of Data and do not fully re ect the structure of the repository network formed by instance-level links (e.g., such important repositories as Freebase, RKBExplorer, and LinkedMDB are not covered). Given the abundance of existing instance-level links, a bottom-up process where the correspondences between classes are captured based on the links between sets of their instances becomes a promising approach. We applied light-weight instancebased ontology matching techniques to a snapshot of the Web of Data which was proposed for the Billion Triple Challenge 2009 competition3 and extracted a large-scale network of ontology mappings. This network provides interesting insights into the use of ontologies on the Web of Data and can be employed to facilitate data integration.
The rest of the paper is organised as follows: in section 2 we brie y outline the ontology matching process we used to extract the mappings and discuss our observations about its applicability and limitations. Then, in section 3 we describe the resulting network of schema mappings we obtained. In section 4 we
overview relevant existing work. Finally, section 5 discusses the limitations of our work and directions for the future work. 2
The snapshot of the Web of Data which we used in our work was proposed for the Billion Triple Challenge 2009 competition4. This is a large-scale dataset containing about 1.14 billion statements. It contains the core portion of the repositories published within the Linking Open Data (LOD) initiative, as well as many smaller datasets retrieved using Semantic Web search engines, such as Watson and Falcon-S. The LOD datasets included into the BTC repository such as DBPedia, Freebase, Bio2RDF, RKBExplorer, Geonames, and others still constitute the core of the Web of Data cloud and are commonly used to connect other datasets. Thus, their schema ontologies are particularly interesting for potential data integration scenarios.
To derive the sets of mappings between these ontologies, we applied a
lightweight matching technique which computes the similarity between a pair of
classes based on the degree of overlap between their instance sets. Originally, we
used this approach to produce schema-level mappings in order to facilitate
further instance coreference resolution and discover previously missing links [
When two classes share at least one individual, we say that there is an overlap relation between these classes. There are two common cases where an individual becomes assigned to several classes de ned in di erent ontologies: { Declared coreference association. In this case, two individuals belonging to di erent repositories are declared to be identical and linked via the owl:sameAs property. This creates an overlap relation between the classes to which the instances belong. { Co-typing. In this case the publishers of a repository structure the data using terms of several ontologies. In this way, one individual can be explicitly assigned to several classes from di erent ontologies. One example is DBPedia, which uses Yago and UMBEL ontologies in addition to its native DBPedia ontology.
These two types of overlaps illustrate di erent aspects of the data structure. Declared association-based overlap relations characterise the distribution of data in di erent repositories and correspondences between sets of their individuals. Co-typing-based mappings mostly highlight the choices of data publishers to use speci c vocabularies to annotate their data. To keep this distinction, in
http://gromgull.net/blog/category/semantic-web/billion-triple-challenge/. and this paper we analyse the declared association-based and co-typing-based overlap mappings separately.
In order to generate all overlap relations present in the dataset, we used the following procedure: 1. Extract all rdf:type relations present in the dataset: A(I), where A is a class and I is an instance of this class. 2. For each class A, generate the set of its instances (extension): e(A) = fIjA(I)g. 3. For each pair of classes A and B, generate the co-typing-based overlap set: ecA \ B = fIjA(I); B(I)g. In total, this constituted about 3.6M co-typingbased overlap mappings (we only considered intersections between classes which did not share the same URI namespace) 4. Extract all owl:sameAs relations present in the dataset (sameAs(I1; I2)) and generate their transitive closure. 5. Generate association-based overlap sets: ea(A \ B) = fI1jA(I1); B(I2); sameAs(I1; I2)g (one sameAs relation corresponds to one element in the set). In total, about 1M (992482) association-based overlap mappings were produced.
For association-based overlap sets we distinguished between a direct class link (when their individuals were explicitly stated in the dataset as identical) and an indirect link (when owl:sameAs relations were inferred using transitivity). Indirect mappings occurred, in particular, when two repositories were connected via a third one (e.g., MusicBrainz and Freebase via DBPedia). Both sets of mappings were ltered to remove general-purpose concepts (such as OWL and RDFS terms) and blank nodes. These two sets of mappings constitute the \raw data" which were later analysed to retrieve valid semantic mappings.
In our original work [
In order to capture the optimal parameters for distinguishing valid semantic mappings, in the experiments described in this paper we employed a machine learning approach. To construct a gold standard set, we have randomly selected a set of 6000 mappings (3000 association-based and 3000 co-typing-based ones) and annotated them manually (\strong overlap" relations were assigned based on subjective judgement). In these initial experiments, annotation was done by one person. After that, we used this gold standard set to train a classi cation model which would assign the relation type to any pair of overlapping classes. Our goal was to nd a suitable classi er to distinguish between valid subsumption and equivalence mappings (owl:equivalentClass and rdfs:subClassOf ) and other mappings.
For the classi er, we included the following features: { ns1, ns2 : namespaces of two class URIs A and B respectively. { je(A \ B)j: the size of the set of instances belonging to both classes A and
B. { je(A)j, je(B)j: sizes of instance sets for classes A and B respectively. { (A; B), (B; A), where (X; Y ) = jej(eX(X\Y)j)j { direct (only for declared association-based links): a boolean value equal to true for direct declared association-based mappings and false otherwise. To test the resulting model, we used the standard 10-fold cross-validation mechanism. After testing, we found that the J48 decision tree algorithm was able to achieve the best performance (Table 1), so this learned classi er was then applied to the whole dataset.
The resulting set of mappings was compared against the set of already existing schema-level relations declared in the dataset. We discovered that the majority of overlap mappings were not covered by explicitly de ned axioms. Only 3119 mappings (2162 and 957 for the declared association-based and cotyping-based subsets respectively) were found to be de ned as rdfs:subClassOf and owl:equivalentClass (or could be inferred), which constituted less than 2.6% and 1.4% of the number of mappings selected by the learned classi er in each case. 3
We applied the learned decision tree models (J48) to our two sets of mappings containing declared association-based and co-typing-based overlap mappings. At the next step, we ltered out redundant mappings: when a class A is found to be a subclass of two classes B and Bsuper where B v Bsuper and the distance metrics are equal ( (A; B) = (A; Bsuper)), then only the mapping A v B remains, and the mapping A v Bsuper is removed. Two resulting sets of mappings were then Property Number of nodes Number of edges cMoanxniemctuimonsnupmerbnerodoef 5301 nNuomdebewritohf ctohnenmecatxioimnsum geonames:Feature cAovnenreacgteionnusmpbeerrnoofde 8.09
Declared association-based Co-typing-based 20365 35578 82422 67620 used to construct networks connecting classes from di erent ontologies. The characteristics of this network are discussed in section 3.1. Then, in order to study the relations between whole vocabularies, we used the original mappings between classes to generate a set of mapping-based links between ontologies. This stage is described in section 3.2. 3.1
We obtained two graphs where classes played the role of nodes and mappings represented edges. The properties of these resulting networks of classes are given in Table 2. To give an overview of the most important \hub" nodes in the network, Table 3 lists the top 10 classes ranked by the number of connections they are involved in.
We can see that the \hub" nodes represent classes representing popular concepts and de ned at the high level in the class hierarchy. Large number of mappings per class is mostly caused by many rdfs:subClassOf relations. After analysing the distribution of mappings per class, we found that in both cases it follows the power law and most classes had only one mapping to another class.
The declared association-based network derived from owl:sameAs links between instances is more connected: average number of mappings per class is 8.09 compared to 3.8 in the co-typing-based case despite the fact that it contains less nodes. This is possibly caused by the \data-level focus" of the LOD initiative: the priority for a data repository owner is to generate instance-level links to other repositories rather than reuse several di erent vocabularies for data description. In this case, class-level mappings automatically derived from owl:sameAs links can be particularly helpful for data integration tasks, because they add new information which was not explicitly stated in any one repository. On the other hand, the co-typing-based network illustrates the impact of ontology popularity: although the graph has more nodes, it is less connected, and a single class foaf:Person contributes to more than 25% of all mappings. From the results we obtained, we can see the strong in uence of DBPedia on the resulting mappings. In the association-based set, 7 out of the top 10 nodes relate to 18137 foaf:Person 3.80 Rank Declared association-based Co-typing-based Rank Name Edges Name Edges 1 geonames:Feature 5301 foaf:Person 18137 2 freebase:people.person 2318 umbel:Person 4533 3 yago:PhysicalEntity100001930 2230 dbpedia:Person 2478 4 yago:Object100002684 2076 foaf:OnlineAccount 1983 5 yago:Abstraction100002137 1759 dbpedia:FootballPlayer 1300 6 yago:Whole100003553 1511 wordnet:Person 1237 7 linkedmdb: lm 1085 dbpedia:Album 996 8 yago:LivingThing100004258 975 dbpedia:Species 920 9 yago:Organism100004475 974 dbpedia:Artist 900 10 yago:CausalAgent100007347 956 dbpedia:MusicalArtist 853 top-level entities from the YAGO ontology. High positions of geonames:Feature and freebase:people.person are also largely due to the number of DBPedia and YAGO classes modelling the respective topics. In the co-typing-based network, we can see the strong presence of the FOAF and WordNet ontologies (largely due to their high reuse in small-scale datasets even before the start of the LOD initiative). Beyond that, all top nodes in the network were produced based on DBPedia instances annotated using di erent schemas. It is interesting to see the high position of the class dbpedia:FootballPlayer. The main reason for it is the large number of YAGO classes (Wikipedia categories) describing this topic.
When we merged two mapping sets into one, we found that only a small subset of mappings (3591) was shared between two networks. Two types of evidence we used produced complementary sets of mappings rather than duplicated each other. 3.2
In order to capture the relations between di erent vocabularies used on the
Web of Data, we generated a set of mapping-based links between ontologies.
In accordance with [
The graphs constructed using declared association-based and co-typing-based evidence are shown in Fig. 1 and Fig. 2. In the declared association-based graph (Fig. 1), the main factor which in uences the position of an ontology in the graph is topic coverage. The top 5 \hub" ontologies with wide coverage do not have a large di erence in the number of connections: YAGO (29), Freebase (28),
UMBEL(27), OpenCYC (26), and DBPedia (23). The 6th and the 7th ranking nodes (LinkedMDB and eurostat), which cover speci c domains, have only 13 connections each. It is interesting to note that although Freebase is connected to less repositories than DBPedia in the LOD cloud5, this does not have an impact at the schema level. This is the e ect of indirect owl:sameAs mappings inferred by transitivity. Connections of domain-speci c ontologies (such as Music ontology or Geonames) point to other ontologies covering the same domain, and, indirectly, to the underlying repositories which contain relevant data. This makes them good starting points when the task is to nd several datasets relevant to a speci c topic. Both networks contain several disjoint subgraphs (5 and 35 respectively), and in both cases the same pattern occurs: there exists one large central cluster including the majority of nodes and several small ones usually including a pair of ontologies (e.g., a cluster fhttp://purl.uniprot.org/core/, http://bio2rdf.org/ns/uniprot#g). In Fig. 1, similarly to the data-level LOD cloud, we can also observe the existence of two \communities" centered around DBPedia and RKBExplorer. At the schema level these are centered around YAGO and AKT ontologies. Both communities are connected via the FOAF ontology (rdfs:subClassOf relations with the foaf:Person class). At the data level, RKBExplorer and DBPedia are connected via two other repositories: DBLP Hannover and DBLP Berlin. The reason for missing schema-level links between AKT and the ontologies used in DBPedia was the omission of intermediate owl:sameAs links on this route, which did not allow indirect declared associationbased class mappings to be produced.
The co-typing-based network (Fig. 2) is substantially larger (746 nodes vs 53) and mainly connects ontologies used outside the LOD cloud (including even legacy schemas like DAML-OIL). In this graph, the distribution of nodes primar
ily illustrates ontology popularity : FOAF (504 connections) and Wordnet (296) get the most connections because they are reused in many datasets. 4
Originally, schema matching approaches in the database and Semantic Web
domains primarily focused on the task of matching two input schemas in isolation
from others [
Recently, with the growing number of public repositories storing data about
overlapping domains, it became important to analyse the emerging network
of interconnections as a whole. The idMesh system[
As mentioned in section 1, schema-level mappings can become a valuable asset for the data publisher who wants to integrate a new repository into the Linked Data environment: for example, having a new repository about music described using the Music ontology, the pool of potential data sources to connect to would include other datasets using the the same ontology, but also repositories which use ontologies mapped to the it (DBPedia, Freebase, LinkedMDB, etc.). From this pool the publisher can select the most comprehensive data source for her needs.
We consider the work described in this paper as our starting point in studying the emerging relations between ontologies on the Web of Data. There are several interesting future directions of research. First, our approach focused on establishing mappings between classes while ignoring mappings between properties, which are equally important in data integration scenarios. Mappings between properties are needed to represent data from di erent ontologies in a uniform
way, which is necessary for applying coreference resolution tools or, in a more general scenario, to present query results to the user.
Second, in the context of our intended scenario (assisting the publisher in the choice of appropriate points of linkage) the quality of mappings had relatively low importance: a mapping is still useful if it connects two classes with a strong degree of overlap, but no strict logical relation holds. This allowed us to use very simple matching techniques to generate schema-level mappings. However, this assumption does not hold for many actual data integration scenarios: in general, a precise SPARQL query is not expected to return irrelevant results. Thus, applying state-of-the-art ontology matching tools to discover high-quality schema mappings in the Linked Data environment constitutes the second direction for future work. 6
Part of this research has been funded under the EC 7th Framework Programme, in the context of the SmartProducts project (231204). The authors would like to thank Paul Groth and Cristophe Gueret for providing the 4store Amazon EC2 community server hosting the BTC dataset.