In this paper, we introduce the vocabulary recommendation system Datavore (Data vocabulary recommender). The tool is oriented towards metadata designers providing ranked lists of vocabulary terms to reuse in the web of data modeling process, together with additional metadata and cross-terms relations. Datavore relies on the Linked Open Vocabulary ecosystem for acquiring vocabularies and metadata. The system is able to deal with noisy and multilingual input data.
Following the fourth principle of linked data, a large number of datasets from di erent domains and using di erent vocabularies have been published and interlinked on the web. To guide data discovery and reuse, catalogs of linked data, such as the Data Hub1, have been created.
With the increasing use of Linked Open Data (LOD ), it becomes more and
more important for data providers not only to publish their data but also to
model and describe them following the LOD best practices2. Here, we draw the
reader's attention to the recommendation of building on, instead of replicating,
existing RDF schema and vocabularies, in an e ort to improve
interoperability [
In this paper, we introduce Datavore, a vocabulary recommender system,
which uses the Linked Open Vocabularies3 (LOV) as a vocabulary search
engine. In addition to a list of ranked recommended concepts/properties, the tool
provides important additional metadata, as well as cross-terms relations in the
form of a set of triples combinations, handling noisy and multilingual input. To
our knowledge, there is only one comparable tool Karma [
We proceed to describe the work ow of the system, shown in Figure 1. Source Terms Extraction (1). The input of Datavore is a list of terms extracted from the data source. In the current version, we parse a loaded CSV le as input and extract the list of column names4. The use of other kinds of structured or semi-structured input data is envisageable.
Cleaning and Translating (2). In most cases, the extracted string of characters needs to be cleaned-up by removing or modifying the unwanted characters. We use the Microsoft Translator java API5 in order to clean up the initial string and render it in a linguistically correct form. For example, translate(CatA~ c gorie, fr, fr) returns Categorie. In case there are no su cient or satisfactory results by using the source language, the system uses the same service to translate the source item into English, the most common language on the LOD. Terms Search (3). We opted for the LOV as a vocabulary search engine, which, to the best of our knowledge, is the only purpose-built vocabulary search engine available on the Web with an up-to-date index. As a design decision, Datavore 4 We note that considering the actual values from the CSV misled the recommendation in most cases that is why we limited ourselves to the column names. 5 https://code.google.com/p/microsoft-translator-java-api/
Authors Suppressed Due to Excessive Length queries the LOV search service with the extracted cleaned or/and translated terms. The result is a list of concepts for each source term ranked by the LOV metric, which is based on the vocabulary terms popularity in the LOD datasets and in the LOV ecosystem.
Metadata Extraction (4). Metadata designers are recommended to select popular
vocabularies found in the search phase but it is not straightforward to judge
which vocabulary is better suited to the application. Datavore queries the LOV
endpoint (/dump le) to extract the needed metadata to help designers to choose
the appropriate vocabularies. As a result, for each concept c, extract: (i) the set
of object properties having c as domain that includes labels and hierarchical
relations, (ii) the set of datatype properties that can be used with c as domain,
and (iii) a link to the vocabulary web site. In addition, we provide a
"test-yourluck " option, which recommends to the user only one, the top ranked, datatype
property. This "lucky" property has the highest Levenshtein string similarity [
Inter-Concept Relations (5). From the extracted lists of recommended concepts, Datavore queries the LOV endpoint (/dump le) to retrieve cross-lists relations, i.e., relations between concepts from di erent lists. These metadata are crucial for selecting the best combination of predicate names to reuse for the input dataset. 3
Imagine a designer who wants to model the data in Table 1 using an ontology editor. Datavore will guide him/her to nd vocabularies to reuse, returning a sorted list of concepts for each column name. For the column "City Name" Datavore queries the LOV using the keyword "City" and returns the sorted list of concepts f"akt:City", "place:City", "lgdo:City", etc.g. When the designer selects the concept "akt:City", Datavore presents to him/her the following metadata: (i) literals (like rdfs:label, rdfs:comment, etc.) and the hierarchical relations of "akt:City", (ii) a set of datatype properties like "foaf:name" that have "akt:City" as rdf:domain to represent the column "City Name". After the concepts extraction, Datavore queries the LOV again to extract inter-columns triples and recommends to the user a set of relations between column names. In our example, the recommended relation between the two columns "Person Name" and "PostalAdress" is the triple: < f oaf : P erson >< akt : hasAddress >< akt : P ostalAddress >. 4
For a proper use of Datavore, we take note of the following.
| Datavore is meant to be used in complementarity with ontology development tools for dataset modeling. id Person Name Profession Lab City Name Postal Adress Country 1 M. Ben Elle PhD Student LIRMM Montpellier 34090 France 2 K. Todorov Assoc. Pr. LIRMM Montpellier 34000 France 3 Z. Bellahsene Pr. LIRMM Montpellier 34000 France
| Modeling Linked Data generally requires an ontology engineer and a
domain expert [
| Datavore uses the LOV search service, for which, as for any search engine, the choice of input keywords is crucial. For that reason, Datavore enables the user to update the source terms from the interface.
| The complexity of the inter-concept relations step is of order O(N 2M1M2), where, N is the cardinality of the source terms, M1 and M2 are respectively the cardinalities of the two compared lists of recommended concepts.
| Datavore has been implemented in Java and it is available as a GU I desktop application6 together with a demonstration video7. 5
We introduced Datavore, a vocabulary recommender system based on LOV assisting linked data modeling. Among the original features of the tool is the fact that it provides metadata of the recommended predicates and that it automatically retrieves existent relations between the predicates to guide metadata design. LOV is a trustworthy search engine but the actual version contains less than 480 vocabularies. In the future, we intend to extend Datavore via the Swoogle search engine, which contains over 10; 000 ontologies.
Acknowledgements. This research is funded under the Datalyse project8.