With more data repositories constantly being published on the Web, choosing appropriate data sources to interlink with newly published datasets becomes a non-trivial problem. While catalogs of data repositories and meta-level descriptors such as VoiD provide valuable information to take these decisions, more detailed information about the instances included into repositories is often required to assess the relevance of datasets and the part of the dataset to link to. However, retrieving and processing such information for a potentially large number of datasets is practically unfeasible. In this paper, we examine how using an existing semantic web index can help identifying candidate datasets for linking. We further apply ontology schema matching techniques to rank these candidate datasets and extract the sub-dataset to use for linking, in the form of classes with instances more likely to match the ones of the local dataset.
H.3.3 [Information Systems]: Information Storage and RetrievalInformation Search and Retrieval Data fusion, data linking, linked data
The fourth principle of Linked Data1 recommends to include links to other URIs so that more information can be obtained by following the links. In order to do that, data publishers must be aware of other repositories containing relevant data and be able to nd existing resources which can be reused or linked to. With the growing number of repositories published within the Linked Data initiative, identifying such datasets and resources can become problematic. As a result, data publishers usually only link their datasets to the popular repositories (such as DBPedia2 and Geonames3). This may not always be the optimal solution in some cases, for example: 1http://www.w3.org/DesignIssues/LinkedData 2http://dbpedia.org 3http://www.geonames.org/ 4http://dblp.l3s.de/ 5http://www.ncbi.nlm.nih.gov/pubmed/ 6http://ckan.net/ 7http://semanticweb.org/wiki/VoiD 8http://rae2001.rkbexplorer.com/ 9http://lod.openlinksw.com/ more detail. Section 4 discusses the results of the initial experiments we performed to test our algorithm. Finally, section 5 concludes the paper.
The problem of determining a set of relevant repositories is a generic one and can occur in di erent contexts. One of the tasks within the SmartProducts project10 involves reusing the data from external semantic repositories to build knowledge bases for smart consumer devices: e.g., to extend the core domain knowledge base of food recipes for a smart kitchen with nutritional data, alternative recipes, health pro les of food products, etc. In order to extend the core domain knowledge base, the developer has to be able to nd relevant repositories on the Web of Data and interlink them with this core knowledge base.
In another scenario, the data.open.ac.uk repository11 aims at publishing various data related to the activities of The Open University (OU)12 according to Linked Data principles. These datasets include, among others, the publications originated by OU researchers, courses provided by the university, etc. Many entities referenced in these datasets are also mentioned in other public repositories. Thus, in order to facilitate data integration, it makes sense to create links from instances used in the data.open.ac.uk datasets to external semantic data stores. Given the range of categories to which data instances belong, it is di cult to select a single external source to link to: e.g., publication venues can be linked to di erent subsets of RKBExplorer, DBLP, PubMed, DBPedia, or Freebase. Moreover, the repository is constantly extended with more instance data for existing topics (e.g., as more research output is published with time) as well as with more topics (as more internal datasets are released online). Selecting relevant sources for linking and selecting speci c individuals to link to within these sources becomes a time-consuming procedure, which needs to be automated as much as possible.
There are several factors which can guide the selection of the repository for linking, in particular:
Degree of overlap. In order to maximise the possibility to reuse external descriptions, the sources which contains more references to the entities stored in the newly published repository are preferable.
Additional information provided by the source. When selecting a source to link to, it is important to take into account how much additional information about entities is provided by each external source: i.e., what properties and relations are used to describe these entities.
Popularity of the source. Linking to URIs de ned in a popular data source or reusing them makes it easier for external developers to nd the published data and use them.
Among these factors, only the degree of overlap heavily relies on instance-level data stored in external repositories. The level of detail of instance descriptions can be obtained from 10http://www.smartproducts-project.eu 11http://data.open.ac.uk 12http://www.open.ac.uk the domain ontology used by the external dataset and, possibly, a few example instances, while the popularity of the source can be estimated based on VoiD linkset descriptors. Therefore, when designing our algorithm, we primarily focused on estimating the degree of overlap between the internal dataset prepared for publishing and potentially relevant external datasets. 3.
The task of nding relevant repositories assumes that there
is a dataset to be published Dp = fOp; Ipg containing a set
of individuals Ip structured using the ontology Op. Each
individual belongs to at least one class c de ned in Op:
I = fijjc (ij); c 2 Opg. On the Web there is a set of Linked
Data repositories fD1; : : : ; Dng such that Dj = fOj; Ijg.
There is a subset of these repositories fD1; : : : ; Dmg which
overlap with Dp, i.e., 8(j m)9(IjO Ij) :
IjO = fikjequiv(ik; ip); ij 2 Ij; ip 2 Ipg, where equiv denotes
the relation of equivalence between individuals. The
meaning of the equivalence relation here depends on the
intentions of the data publisher and the type of links (s)he wants
to generate: e.g., owl:sameAs links or direct reuse of URIs
assume that URIs must be strictly interchangeable while
rdfs:seeAlso may only assume some kind of similarity (see [
Since the actual discovery of links is usually performed by
an automated tool (such as Silk [
The main obstacle with these tasks is the need to identify IO the overlapping subset of instances j j j from each external dataset. Downloading whole datasets or applying data linking tools to their complete sets of instances is often unfeasible due to their size and required computational time, network load, and local disk space. Thus, the degree of overlap has to be estimated, and keyword search services can be utilised to perform this task. 3.1
We assume that a semantic keyword search service takes
as its input a set of keywords K = fk1; : : : ; kig. As output, it
returns a set of potentially relevant individuals which may
belong to di erent repositories: Ires = I1res [ I2res
Imresg, where Ijres Ij. For returned individuals ijk [2 :I:jr:es[,
their types fcjk jcjk (ijk)g are also available in the search
results. An example of the search service which satis es this
assumption is Sig.ma [
In order to nd potentially relevant individuals for individuals from the newly published dataset Dp, we query the search service using the labels of individuals (values of rdfs:label, foaf:name, dc:title, etc.) as keywords. Then, these query results are aggregated to estimate the degree of overlap of di erent data sources (Fig. 1). The procedure consists of the following steps: 1. Randomly selecting a subset of individuals from Dp belonging to a class cp. This is done in order to reduce the number of queries to the search service in case where the complete extension set of individuals is too large. On the other hand, the subset must be large enough to produce reliable ranking of sources. 2. Querying the search service (Sig.ma) for labels of each individual in the selected subset. The results of each search are returned as an RDF document, which includes the references to individuals, their sources, and the classes they belong to. 3. Aggregation of the search results. RDF documents returned by Sig.ma are loaded into a common repository, and the individuals ijk are grouped according to their sources Dj. 4. Data sources are ranked according to the number of their individuals returned by the search service jfijkjijk 2 Djgj.
In our approach we assume that the relevance function used by the search service to select query answers serves as an approximation of the identity function equiv(). In the general case, this is in not true due to ambiguity of labels and the fact that search services may not always achieve 100% precision. Taking a su ciently large subset of individuals to search makes it possible to reduce the impact of \false positives" returned by the search engine.
After applying these steps to our test scenarios (see section 4), we found that the rankings obtained using this procedure are still likely to be imprecise for two main reasons: Inclusion of irrelevant sources. For individuals belonging to classes with highly ambiguous labels, many \false positives" in the set of answers can result in irrelevant repositories achieving high ranking positions. For instance, when searching for speci c subcategories of people, any source mentioning su ciently large number of people would be considered relevant: e.g., Twitter and DBLP were highly ranked when searching for music contributors.
Inclusion of irrelevant classes. Resulting sets often contained classes which would not allow selecting appropriate candidate individuals by a matching tool. Sometimes a generic superclass was ranked higher than the correct class: e.g., dbpedia:Person was ranked higher than a more relevant dbpedia:MusicalArtist. In other cases, completely irrelevant classes were included: e.g., for scienti c journals the class akt:Publication-Reference describing speci c volumes of journals was ranked higher than akt:Journal.
In order to overcome these issues, our approach includes the second stage: ltering of search results using ontology matching techniques. 3.2
In order to lter out irrelevant search results, our approach
can utilise mappings between classes provided by existing
schema matching tools (Fig. 2). In our experiments we
utilised ontology mappings produced by two algorithms:
CIDER [
As the rst step of the ltering procedure, CIDER is applied to measure similarity between the class cp in Dp, for which overlapping sources have to be found, and each of the classes cjk appearing in the aggregated search results. Then, a threshold is applied to lter out classes with low similarity scores. Remaining classes from the search results constitute the set of \con rmed" classes Cconfirmed. At the next stage, this set of \con rmed" classes is enriched using the mappings obtained using instance-based matching. For 13http://vmlion25.deri.ie/ each class ci 2 Cconfirmed, all mappings from the BTCbased set where cA v ci are selected, and all cA are added into Cconfirmed. Then, the resulting set of search results is ltered so that only individuals belonging to \con rmed" classes remain. In our tests described in section 4, the ltering stage led to improved precision in the resulting ranking.
In our initial tests, we have applied the approach described in section 3 to the following datasets:
ORO journals. A set of 3110 journals mentioned in the ORO repository constituting a part of data.open.ac.uk.
Each individual belongs to the class bibo:Journal 14.
LinkedMDB lms. A subset of 400 randomly selected instances of the class movie: lm15 representing movies in the LinkedMDB repository.
LinkedMDB music contributors. A subset of 400 randomly selected instances of the class movie:music contributor representing music contributors for lms in the Linked
MDB repository.
For each individual in these sets, we queried Sig.ma using their labels as keywords. First, we produced the ranking of sources using the whole set of search results returned by Sig.ma as described in section 3.1 and counted the number of actually relevant data sources among the top-10 ranked ones. Then, we applied the ltering mechanism using ontology schema matching results and checked the relevance of remaining sources. The results we obtained are presented in Table 1: for each dataset it shows the list of top ranked sources as well as our judgement whether these sources were actually relevant (column \+/-"). In the table, \(RKB)" denotes the datasets from RKBExplorer and \open EAN" corresponds to openean.kaufkauf.net. The \+/-" value denotes that the source could only be considered relevant in a speci c context. In particular, the repositories listing lm DVDs as trade commodities are relevant in the context of e-commerce, but not, e.g., as reference sources for students. For both LinkedMDB datasets, we did not consider the LinkedMDB repository itself when it was returned in the search results. As we can see from the results, the initial search-based ranking managed to discover relevant datasets for the sets of individuals in question. Top-ranked sources in the Journals and Films categories contained relevant individuals which could be linked to the individuals in Dp, and their sets of individuals are to a large degree overlapping. For music contributors, the proportion of irrelevant sources was substantially larger due to higher ambiguity of human names. The ltering stage in all cases resulted in improving the ranking precision: only relevant sources were con rmed. However, if we look at the ranking of ontological classes (Table 2), we can see that correctly identifying classes presents a number of issues. The table shows the highest ranking classes returned after each stage of the algorithm (only one highest-ranking class from each ontology is shown). Top-ranked classes produced from the search results usually represent high-level concepts and correspond to superclasses of the original class: e.g., 14http://purl.org/ontology/bibo/Journal 15http://data.linkedmdb.org/movie/ lm
After ltering
Top-ranked
rae2001(RKB)
DBPedia
dblp.l3s.de
Freebase
DBLP(RKB)
eprints(RKB)
+/+
+
+
+
+
+
DBPedia
Freebase
Freebase
DBPedia
+
+
+
+
After ltering
Top-ranked
akt:Journal
yago:Periodical
swrc:Journal
dbpedia:Work
freebase:book.periodical
dbpedia:Film
yago:Movie
freebase: lm. lm
freebase: lm.
music contributor
yago:American
TelevisionComposers
foaf:Document or dc:BibliographicResource for journals,
dbpedia:Work for movies, and foaf:Person for musicians.
Considering all instances of these classes as candidates for a
data linking tool can lead to several problems. Matching
algorithms usually implement pairwise comparison of
individuals, so choosing all instances of a generic class as
candidates for matching is likely to increase the computational
time substantially. Also, less ne-grained feature selection
is possible because important discriminating properties are
often subclass-speci c, and only properties common for all
subclasses are de ned for top-level concepts. This, in turn,
can lead to lower quality of produced links, in particular,
greater number of \false positives" [
The ltering stage largely removed these problems so that only classes with a stronger degree of semantic similarity 6. were con rmed. However, it also reduced the recall in cases where a directly corresponding class was not present in the external ontology: e.g., individuals from dotac.rkbexplorer.com and oai.rkbexplorer.com, which only used the generic class dc:BibliographicResource were not considered as relevant sources 7. for linking journals. Similarly, many relevant classes were ltered out because they were not considered as exact matches or subclasses of the class movie:music contributor (e.g., mo:MusicArtist and dbpedia:MusicalArtist ).
Identifying relevant sources for interlinking already can
present a non-trivial problem, and in future this issue is
likely to become more important. The Linked Data cloud is
constantly growing, and in order to make its use widespread,
data owners must be able to publish their datasets without
extensive knowledge about the state of the Web of Data
or assistance from the research community. Interlinking is
an important part of the publishing process and the one
which can require substantial exploratory work with
external data. Thus, this process has to become straightforward
for data publishers and, preferably, require minimal human
involvement. While the problem of link discovery has been
addressed by several approaches (see, e.g., SILK [
Integration into the generic data publishing work ow in order to provide a structured approach for data publishing within the organisation.
Improvement of the search quality, in particular, the
ltering stage. One particular route involves the
analysis of possible choices of relevant sources and classes by
estimating potential loss of precision and recall (e.g.,
see [
Another potentially interesting research direction is related to the development of semantic indexes. Search for relevant data repositories can become a novel interesting use case in addition to the more common search for entities and documents. In order to support it, new types of search services can be valuable: for example, batch search for a large array of resource labels instead of multiple queries for small sets of keywords, which increase number of server requests and overall processing time.
This research has been partially funded under the EC 7th Framework Programme, in the context of the SmartProducts project (231204).