Mapping RDF data to ArangoDB Graphs

This document serves as a guide to getting started with RDF and ArangoDB. In it we highlight some potential approaches for working with RDF graphs and then some of the considerations when attempting to bring RDF data into ArangoDB. The concept of working with RDF graphs in ArangoDB is not fully supported, but we would like to start introducing solutions for mapping and start the conversation with the community. We welcome any feedback from the community on ways to improve the ingestion of RDF graphs as we work on an official implementation for ArangoDB. You can accomplish a lot with ArangoDB and RDF with a few workarounds, depending on your needs but be sure to see the caveats section for some considerations.

Getting Started

RDF graphs are purely directed graphs with no properties on vertices or edges. In RDF, everything is referenced by edges; these edges are known as statements. Statements are in the form [subject, predicate, object, [graph]]. Statements pose an interesting challenge when trying to interpolate an RDF graph as a property graph. ArangoDB is uniquely suited to handle this due to having edges which have a similar structure to vertices and thus can be used to resemble an RDF graph or to instead represent RDF statements as properties associated with a subject. For this discussion, it is helpful to have an example; the following is a simple RDF graph for Sir Arthur Conan Doyle.

<?xml version="1.0" encoding="UTF-8"?>
<rdf:RDF
   xmlns:rdf="http://www.w3.org/1999/02/22-rdf-syntax-ns#"
>
  <rdf:Description rdf:about="http://dbpedia.org/resource/Arthur_Conan_Doyle">
    <rdf:type rdf:resource="http://dbpedia.org/ontology/Person"/>
    <rdf:type rdf:resource="http://dbpedia.org/ontology/Writer"/>
    <rdf:type rdf:resource="http://www.w3.org/2002/07/owl#Thing"/>
    <rdf:type rdf:resource="http://dbpedia.org/ontology/Artist"/>
  </rdf:Description>
</rdf:RDF>

The above code is RDF in RDF/XML format. This syntax can be considered the original/standard RDF syntax. However, many different serializations exist, and thanks to their readability and ease of parsing, has grown in popularity.

<http://dbpedia.org/resource/Arthur_Conan_Doyle> <http://www.w3.org/1999/02/22-rdf-syntax-ns#type> <http://www.w3.org/2002/07/owl#Thing> .
<http://dbpedia.org/resource/Arthur_Conan_Doyle> <http://www.w3.org/1999/02/22-rdf-syntax-ns#type> <http://dbpedia.org/ontology/Person> .
<http://dbpedia.org/resource/Arthur_Conan_Doyle> <http://www.w3.org/1999/02/22-rdf-syntax-ns#type> <http://dbpedia.org/ontology/Artist> .
<http://dbpedia.org/resource/Arthur_Conan_Doyle> <http://www.w3.org/1999/02/22-rdf-syntax-ns#type> <http://dbpedia.org/ontology/Writer> .

This is the same RDF graph but serialized to N-Triples format. This format represents the same data from the same source. While the exploration of RDF, its syntax, and various serializations are beyond the scope of this guide, it is important to consider that the data ingestion isn’t always in the same format when dealing with RDF and is one of the few challenges we encounter when dealing with RDF graphs.

The following is an example of how this RDF graph could be modeled in ArangoDB. This approach:

• Maps the subject and object as vertices
• Maps the predicate as the connecting edge
• Hashes the referent to the _key value
• Preserves the full referent as a text property

There can be issues that arise with this simple approach, and we discuss them throughout the guide, but this is a good place to start as it shows the potential for mapping the data.

Generate this graph for yourself by running the notebook.

Considerations

Literals

In RDF, even literal values are referenced by edges. Literals cannot have outgoing edges (i.e., cannot be the subject of a statement). RDF uses the XSD type system for literals, so the string “Fred” is represented as "Fred"^^xsd:String or fully expanded as "Fred" ^^http://…". Literals can also contain language and locale tags, for example, "cat@en" ^^xsd:String and "chat@fr"^^xsd:String. These language tags can be useful and would ideally be preserved.

Literals could be added as a property instead of creating a separate vertex; this takes better advantage of using a property graph. If you are coming from a triple store or downloading your data using a SPARQL query you could handle these properties when exporting.

Internationalized Resource Identifiers (IRIs)

Prefixes

In RDF, it is common to use namespace prefixes with references for ease of parsing. This can be easily handled with a property graph in a few ways. The easiest approach is to add the statement prefixes to the document. This keeps the prefixes close to the data but results in a lot of duplicated fields. Another approach would be to append the prefix and form the full URI as a property.

Identifiers

IRIs (ex: http://dbpedia.org/resource/) are used as universal identifiers in RDF but contain contain special characters, namely : and /, which make them not suitable for use as an ArangoDB _key attribute. This is the reason the previous example hashes the IRI value. This has a downside of relying on the hashing algorithm and in our case MD5 is one way so it becomes required to store the full IRI string.

Blank Nodes

Blank nodes are identifiers that have local scope and cannot (must not) be referenced externally. Blank nodes are usually used in structures like lists and other situations where it is inconvenient to create IRIs. They will cause problems when reconciling differences between graphs. Hashing these values as well is a way to work around them but as they are considered temporary identifiers in the RDF world they could pose consistency issues in your RDF graph.

Serializations

There are numerous RDF serializations, including XML and JSON based serializations and gzipped serializations. Third party libraries will likely handle all of the serializations but it is a step that may effect how data is imported.

Ontology, Taxonomy, Class Inheritance, and RDFS

The final consideration is something that for many is the core of RDF and semantic data: Ontologies. Not just ontologies but also class inheritance, and schema validation. One method would be add the ontology in a similar way to what has been suggested for the RDF graphs as ontologies are usually structured in the same way (or can be). However, how do you verify your data is always referencing the classes in your ontology? How can this data be used to make inferences if it cannot be validated? ArangoDB has schema validation but this is limited to JSON schema and does not complex class checking.

One approach to this would be to use a Foxx microservice to serve multiple knowledge graph functions.

DIY with Foxx

A Foxx service could be used to perform lookups against a document collection containing schema-like requirements. The data scientist could use this Foxx service as the ingestion entry point, meaning the data import would need to be validated before entry (resulting in slow imports).

Additionally, the Foxx service could instead be used as an eventual consistency checker that evaluates the data after import and either produces errors or take some other action. This could be extended to offer inference and rule generation based on the data inserted. Most of this functionality would rely on the Foxx queues feature or require manual intervention.

Using Foxx has its own drawbacks as it requires development efforts, consumes resources on the DB-Servers and Coordinators, and uses node which isn’t an option for all organizations.

The benefits of Foxx are the flexibility to program the precise needs of an organization and with the potential for high performance. The service would be located close to the data and has c++ access as a first class citizen. This has the potential to reduce any negative performance impact that might normally come with abstracting away this functionality.

RDF-star

While there is still work to be done when trying to bring RDF graphs into property graphs, there is one approach that aims to make it more worthwhile. The RDF-star implementation aims to bring the gap between RDF graph and property graph. Our initial internal approach to bring RDF data into ArangoDB will be to take advantage of the RDF-star specification. The RDF-star specification allows for nesting attributes in RDF statements to more closely mirror the benefits of a property graph. It is still in draft form but multiple vendors and libraries have already added support for it. If you haven’t already give the specification a look and let us know if you would like to see it in ArangoDB.