Most of the data on the Web is stored in relational databases. In order to make the Semantic Web grow we need to provide easy-to-use tools to convert those databases into linked data, so that even people with little knowledge of the semantic web can use them. Some programs able to convert relational databases into RDF ¯les have been developed, but the user still has to link manually the database attribute names to existing ontology properties and this generated \linked data" is not actually linked with external relevant data. We propose here a method to associate automatically attribute names to existing ontology entities in order to complete the automation of the conversion of databases. We also present a way - rather basic, but with low error rate - to add links automatically to relevant data from other data sets.
Even though signi¯cant research and development e®orts have been made, the
achievement of the vision of the Semantic Web remains remote. The amount of data on the
Semantic Web remains marginal in comparison with the traditional Web. The
importance of revealing relational data and making it available as RDF and as Linked Data[
In this paper, we will present a method to generate automatically the mapping
between attribute names and existing ontology entities, completed with a method to
add links automatically to external data. Then, we will present the application of this
method on relational databases applied on the D2RQ Mapping system and the D2R
Server[
A presentation of our software AuReLi (Automatic Relational Database to Linked
Data Converter) can be found at http://ri-www.nii.ac.jp/AuReLi/
is on one side a database and on the other side we have several ontology descriptions.
Thus, structure-based method are not relevant here. We are seeking to compare the
name of an attribute in a database with the name of an ontology property. These
names can be composed by one or several words: the ¯rst step is to split the name into
a set of words in order to compare the words of each set. The success of the matching
depends on the correctness of the word decomposition. The words composing names
of ontology properties and of attributes in a database are usually either separated by
special characters, for instance product_name, or by a change of case, e.g. ProductName.
As sometimes it is not the case, we completed this simple splitting method with a
method based on the presence of the words in a dictionary such as the WordNet
dictionary1 used here. After doing the previous splitting, it is necessary to check if
the resulting words exist in the dictionary. If that is not the case, then we try to
split it into words that are in the dictionary. However, because it is possible that
the word is not in the dictionary but some part of it is, we will only keep the result
if all the decomposed parts are present in the dictionary. With this method, even
productname will be correctly split. The second step is to compare the resulting set of
words of the attribute name with the sets of words of all the ontology entities, and then
return the best match. In order to compare the words, we use string-based similarity
measures2, especially Jaro-Winkler, and WordNet similarity measures3: Lin[
Once the mapping is done, in order to have true linked data, we want to add links to relevant data. The idea is to make a SPARQL query on a given data set. If you know the target data and its ontology entities, you can speci¯cally build SPARQL queries for this data set to get links. But here, in a more general setting, we do not have this information. However, there is a property common to most of the data sets: rdfs:label. Even better, this property is especially good because it is usually at the same time short and clearly de¯ning the data. Therefore, if the rdfs:label property was correctly set on your data, the SPARQL query based on this property should not return wrong links and has good chances to ¯nd a result if there is a related data in the target data set. 3
We produced a reusable Java library and used the D2RQ Map and the D2R Server[
http://wordnet.princeton.edu/wordnet/download/ 2 S. Chapman: SimMetrics Java library:
http://www.dcs.shef.ac.uk/~sam/simmetrics.html 3 D. Hope: Java WordNet::Similarity:
http://www.cogs.susx.ac.uk/users/drh21/ the user so that he can check and make changes if necessary. It also allows the user to choose which attributes to use as labels for the rdfs:label property.
The D2R Server was also modi¯ed to add links automatically in the generated data. If the feature is activated, it makes a SPARQL query on DBpedia for each request of the user and add the link to the data if there was a result. We used DBpedia because it is currently one of the biggest and the most general linked database. 4
Five databases from di®erent sources, with di®erent size and about di®erent topics were used for the tests: Northwind4, World and Sakila5, Automobile6, World Development Indicator7
There are approximately three hundred attributes in those ¯ve databases: after a manual check of the mappings, 79,66% of the attributes were correctly mapped. The wrong mappings are explained by the fact that some attributes were too speci¯c and consequently could not match any existing ontology property in the ontology descriptions used in the experiment. Another limit is the use of acronyms or short abbreviations, which did not produce a correct mapping either. The generation time was around one minute for each database. The mapping generated automatically can be seen in Fig. 2 for the World database. On the left and the middle are the table names and the attribute names, on the right are the matched ontology entities. We can observe for instance that the attribute GNPOld do not have good corresponding property and thus is mapped with foaf:OnlineAccount which is obviously irrelevant. But on the other hand, Percentage becomes dbpedia:part, which is quite good since a percentage is a part of something. This matching is due to WordNet because it would not have been found by a string-based similarity measure.
For the server, the use of the feature to automatically add links is slightly slowing down each request of the user because it needs the answer of the SPARQL query. It 4 Example database for the Microsoft SQL Server: http://www.microsoft.com/downloads/details.aspx?FamilyID= 06616212-0356-46a0-8da2-eebc53a68034 5 Two example database from the MySQL website:
http://dev.mysql.com/doc/index-other.html 6 Data set from the UCI Machine Learning Repository:
http://archive.ics.uci.edu/ml/datasets.html 7 database from the World Bank Data Catalog:
http://data.worldbank.org/data-catalog becomes problematic if the external data set is slow or do not answer to the query. The results on the rdfs:label property on DBpedia are usually good, providing the labels in the mapping are correct. The principal case where the added links are wrong is in the case of homonyms, e.g. cities such as London, England and London, Canada. 5
The automatic mapping generation is a di±cult problem which renders almost impossible the automatic production of a 100% correct mapping. Nevertheless, even if the user still needs some knowledge of the Semantic Web, we managed to simplify the process with a user-friendly interface where the user only has to check the correctness of the proposed mapping. The automatic addition of links in the generated RDF is simple and functional, and can easily be extended to add a greater variety of links.