RDF-star was recently proposed as a convenient representation to annotate statements in RDF with metadata by introducing the so-called RDF-star triples, bridging the gap between RDF and property graphs. However, even though there are many solutions to generate RDF graphs, there is no systematic approach so far to generate RDFstar graphs from heterogeneous data sources. In this paper, we propose RML-star, an extension of the RML mapping language to generate RDFstar. We introduce the extension of the RML ontology and the associated specification with representative examples. URL: https://w3id.org/kg-construct/rml-star RDF-star was proposed as a compact representation to annotate statements in RDF with metadata [4]. For instance, the following declares that Bob claims Alice was born in 1996: :bob :claims <<:alice :birthYear 1996>>. Following the uptake of the proposed solution, a W3C Community Group was formed3 and a W3C Draft Report [5] was recently released with improvements over the original proposal. By now, several RDF-related programming libraries, e.g., Eclipse RDF4J, Apache Jena, RDF.rb, and N3.js, and RDF graph database systems, e.g., Blazegraph, AnzoGraph, Stardog and GraphDB, have adopted RDF-star4. However, no mapping language supports the generation of RDF-star graphs so far. Most data are still heterogeneous, represented in different formats (e.g., relational databases, CSV, JSON, or XML). One of the most common approaches nowadays to integrate them into RDF graphs is the use of declarative mapping
languages such as R2RML [
In this paper, we propose RML-star, an extension of RML to generate RDFstar graphs from heterogeneous data sources. We introduce a set of new classes and properties that allow describing how RDF-star datasets can be created from heterogeneous data sources in a systematic manner, using the same mapping language as to generate RDF datasets. We also introduce the RML-star specification that explains in detail how these extensions should be used and implemented. 2
The aim of RML-star is to generate RDF-star triples by applying a set of additions and extensions over RML. The changes over the RML vocabulary include two new classes and three object properties, and the modification of one object property (Figure 1). The specification of RML-star with the corresponding ontology is available online at https://w3id.org/kg-construct/rml-star.
Throughout this section, we rely on an example to demonstrate RML-star. We consider two data sources: the CSV file in Listing 1 and the JSON file in Listing 2. The RML-star mapping for these data sources is given in Listings 3 and 4. Finally, in Listing 5 we show the generated RDF-star graph. RML. Before we explain the RML-star extensions, we summarize how an RML mapping is defined. RML consist of a set of Triples Maps which include a Logical Source (lines 2 and 11 of Listing 3) to access the data sources, a Subject Map to generate the subjects of the triples (lines 3-4 and 12-13 of Listing 3), and multiple Predicate-Object Maps to generate the predicates and objects (lines 5-9 and 14-17 of Listing 3). Predicate-Object Maps are in turn composed of Predicate Maps and (Referencing) Object Maps (lines 6 and 7-9 of Listing 3 respectively). A Referencing Object Map uses the Subject Map of another Triples Map to generate the objects. Since a Referencing Object Map may involve two different data sources, join conditions can be specified.
PERSON , BIRTHYEAR, CLAIMER, CONFIDENCE alice , 1996 , bob , 0.9 charlie, 2002 , daniel , 0.3 [ { "PATIENT": "alice",
"HOSPITAL": "Juan Ramon Jimenez" }, { "PATIENT": "charlie", "HOSPITAL": "AZ Maria-Middelares" } ]
source :birthyears (CSV).
Listing 2: Contents of the logical source :hospitalrecords (JSON). :innerTM a rml:NonAssertedTriplesMap ; rml:logicalSource :birthyears ; rml:subjectMap [
rr:template ":{PERSON}" ] ; rr:predicateObjectMap [ rr:predicate :birthYear ; rml:objectMap [ rml:reference "BIRTHYEAR" ; rr:dataType xsd:integer ]] . :outerTM a rr:TriplesMap ; rml:logicalSource :birthyears ; rml:subjectMap [
rr:template ":{CLAIMER}" ] ; rr:predicateObjectMap [ rr:predicate :claims ; rml:objectMap [
rml:embeddedTriplesMap :innerTM ]] .
mapping.
It creates embedded triples that are not asserted. :outerOuterTM a rr:TriplesMap ; rml:logicalSource :birthyears ; rml:subjectMap [
rml:embeddedTriplesMap :outerTM ] ; rr:predicateObjectMap [ rr:predicate :confidence ; rml:objectMap [ rml:reference "CONFIDENCE" ; rr:dataType xsd:float ]] . :joiningTM a rr:TriplesMap ; rml:logicalSource :hospitalrecords ; rml:subjectMap [ rml:embeddedTriplesMap :innerTM ; rr:joinCondition [ rr:child "PATIENT" ; rr:parent "PERSON" ]] ; rr:predicateObjectMap [ rr:predicate :recordedBy ; rml:objectMap [
rml:reference "HOSPITAL" ]] .
4:
Mapping
extension of Listing 3, containing nested triples and multiple data sources. that reason, it belongs to the domain of rml:subjectMap and rml:objectMap properties. The original properties, rr:subjectMap and rr:objectMap had cardinality restrictions that prevent extending them to include Star Map in their domain. These additions are used exactly as the original ones in any other sense.
The object property rml:embeddedTriplesMap connects the Star Map to the Triples Map that defines how the RDF-star triples will be generated. A simple example of a Star Map is shown on lines 16-17 of Listing 3: it embeds triples generated by the Triples Map :innerTM in the objects of the triples generated by the Triples Map :outerTM. This results in the triples shown on lines 1-2 of Listing 5 when given Listing 1 as input.
Non-Asserted Triples Map. An asserted RDF-star triple is a triple that is an element of an RDF-star graph, as opposed to an embedded RDF-star triple, that only appears in the subject or object of another RDF-star triple. In RMLstar, all generated triples are considered by default asserted RDF-star triples. To specify that a generated triple is embedded but not asserted, we introduce the Non-Asserted Triples Map (rml:NonAssertedTriplesMap) as a subclass of Triples Map (rr:TriplesMap). This Triples Map has the same expressiveness as every other Triples Map and just adds the information of being non-asserted. For instance, :innerTM (line 1 of Listing 3) is declared to be a Non-Asserted Triples Map and, as a result, the :birthYear triples it generates are not present in Listing 5 as asserted triples: they only occur as embedded triples.
This structure allows the recursion of Triples Maps to nest as many embedded triples as needed. For example, the Triples Map :outerOuterTM generates triples that have embedded triples generated by :outerTM as their subject, and :outerTM in turn generates triples with embedded triples from :innerTM as their object. As a result, :outerOuterTM generates triples containing two levels of embedded triples (lines 3-4 of Listing 5).
An Embedded Triples Map can generate triples using different data sources. Thus, the Star Map needs to have join conditions to combine such data sources. To achieve this, the property rr:joinCondition is extended to include Star Map in its domain. This property, in contrast to rr:objectMap and rr:subjectMap, is easily extended due to the lack of restrictions in the original vocabulary. On lines 12-16 of Listing 4 a Star Map is declared which joins the data sources in Listings 1 and 2 on equal values of the PERSON column and the PATIENT attribute. It creates the triples on lines 5-6 of Listing 5. 3
In this paper, we present RML-star, an extension of RML, which allows generating RDF-star graphs from heterogeneous data sources. We include a set of new classes and properties while maintaining the general structure of R2RML and RML. With this proposal, we aim at promoting the adoption of RDF-star and pave the way so other mapping languages provide similar extensions. RML-star is discussed within the Knowledge Graph Construction W3C Community Group5 5 https://w3id.org/kg-construct and will be part of the specifications’ suite developed by the group. Thanks to this solution, we devise a promising future work line on the development of efficient and scalable systems to generate RDF-star graphs.
Acknowledgments This research is financially supported by Ministerio de Ciencia e Innovaci´on, Spain, under grant Knowledge Spaces (PID2020-118274RB-I00) and by an FPI grant (BES-2017-082511).