Convert with RML

Use the RDF Mapping Language (RML) to map your structured data (CSV, TSV, XLSX, SPSS, SQL, XML, JSON, YAML) to RDF using a declarative mapping language.

Create mapping for a dataset#

You can run this command at the root of your repository to generate the dataset mappings files in the datasets folder, you will be prompted to enter some metadata about the dataset to create.

d2s new dataset

The dataset readme, mappings, metadata, and download files are created in the datasets/$dataset_id folder. Check the download script generated in datasets/$dataset_id/scripts/download.sh and edit it if needed.

We use bash for it's performance and reliability with large file download. But you are free to use a python script or other documented methods.

Define mappings#

We recommend to use YARRRML, a mapping language to make the definition of RML mappings easier using a simplified YAML, which is then converted to proper RML.

The Matey web UI 🦜 is available to easily write and test RML mappings in YAML files using the YARRRML simplified mapping language. The mappings can be conveniently tested in the browser on a sample of the file to transform.

Recommended workflow to easily create and test RML mappings:

Use the Matey web UI 🦜 to write YARRRML mappings, and test them against a sample of your data
Copy the YARRRML mappings to a file with the extension .yarrr.yml
Copy the RML mappings to a file with same name, and the extension .rml.ttl
Optionally you can automate the execution in a GitHub Actions workflow.

Specifications

RML Specifications can be found as a W3C unofficial draft.
See the rml.io website for more documentation about RML and the various tools built and deployed by Ghent University.

YARRRML package

YARRRML can also be parsed locally or automatically using the yarrrml-parser npm package:

npm i @rmlio/yarrrml-parser -g
yarrrml-parser -i mappings.yarrr.yml

Example of a YARRRML mapping file using the split function on the | character:

prefixes:
  grel: "http://users.ugent.be/~bjdmeest/function/grel.ttl#"
  rdfs: "http://www.w3.org/2000/01/rdf-schema#"
  gn: "http://www.geonames.org/ontology#"

mappings:
  neighbours:
    sources:
      - ['countries.csv~csv']
    s: http://www.geonames.org/ontology#$(ISO)
    po:
      - [a, gn:Country]
      - p: gn:neighbours
        o:
            function: grel:string_split
            parameters:
                - [grel:valueParameter, $(neighbours)]
                - [grel:p_string_sep, "\|"]
            language: en

grel:p_string_sep separators needs to be escaped with \
See the full list of available default functions.
Additional function can be added by integrating them in a .jar file, see the documentation.

Generate nquads

You can also generate nquads by adding the graph infos in the rr:subjectMap in RML mappings (or just g: in YARRRML):

rr:graphMap [ rr:constant <https://w3id.org/d2s/graph> ];

⚠️ Most RML engines does not support YARRRML by default, so you will need to convert it to RML and use the RML mappings for the conversion.

Tools for RML conversion#

There are multiple tools available to generate RDF from RML mappings, with various efficiency, stability, and features.

rmlmapper-java
- Reference implementation, written in java
- Not suited for large files
- Supports custom functions (in java, compiled as separate .jar files)

RMLStreamer
- Streaming implementation for large files, written in Scala
- Works well for really large CSV files
- Can be parallelized on Apache Flink clusters
- Does not support custom functions yet

SDM-RDFizer
- Written in Python
- Can use a separate tool, Dragoman, for executing custom functions

morph-kgc
- Written in Python
- Does not support custom functions

RocketRML
- Written in JavaScript
- Provide an easy way to define custom functions

We currently only implemented the rmlmapper-java and the RMLStreamer in d2s, but you are encouraged to use the tool that fits your needs.

Convert with the RML Mapper#

The rmlmapper-java execute RML mappings to generate RDF Knowledge Graphs.

Not for large files

The RML Mapper loads all data in memory, so be aware when working with big datasets.

Download the rmlmapper .jar file at https://github.com/RMLio/rmlmapper-java/releases
Run the RML mapper:

java -jar rmlmapper.jar -m mapping.ttl -o rdf-output.nt

Run automatically in workflow

The RMLMapper can be easily run in GitHub Actions workflows, checkout the Run workflows page for more details.

- name: Run RML mapper
  uses: vemonet/rmlmapper-java@v4.9.0
  with:
    mapping: mappings.rml.ttl
    output: rdf-output.nt
  env:
    JAVA_OPTS: "-Xmx6g"

Convert with the RML Streamer#

The RMLStreamer is a scalable implementation of the RDF Mapping Language Specifications to generate RDF out of structured input data streams.

Work in progress

The RMLStreamer is still in development, some features such as functions are yet to be implemented.

To run the RMLStreamer you have 2 options:

Start a single node Apache Flink cluster using docker on your machine.
Use the DSRI Apache Flink cluster (especially for really large files).

Documentation

See the DSRI documentation to deploy Apache Flink
Checkout the documentation to convert COHD using the RMLStreamer on the DSRI.

Prepare files#

Copy the RMLStreamer.jar file, your mapping files and data files to the Flink jobmanager pod before running it.

For example:

# get flink pod id
POD_ID=$(oc get pod --selector app=flink --selector component=jobmanager --no-headers -o=custom-columns=NAME:.metadata.name)
DATASET=my-dataset
oc rsh flink-jobmanager-7459cc58f7-5hqjb
oc exec $POD_ID -- mkdir -p /mnt/project
# If script run from datasets/dataset1/scripts/ :
oc cp ../../mappings $POD_ID:/mnt/project/
chmod +x /mnt/project/datasets/$DATASET/scripts/download.sh
oc exec $POD_ID -- /mnt/project/datasets/$DATASET/scripts/download.sh
oc exec $POD_ID -- wget -O /mnt/RMLStreamer.jar https://github.com/RMLio/RMLStreamer/releases/download/v2.0.0/RMLStreamer-2.0.0.jar

Run the RMLStreamer#

Example of command to run the RMLStreamer from the Flink cluster master:

nohup /opt/flink/bin/flink run -p 128 -c io.rml.framework.Main /mnt/RMLStreamer.jar toFile -m /mnt/mappings.rml.ttl -o /mnt/rmlstreamer-mappings-output.nt --job-name "RMLStreamer mappings.rml.ttl" &

Check the progress

The progress of the job can be checked in the Apache Flink web UI.

Merge and compress output#

The ntriples files produced by RMLStreamer in parallel:

cd /mnt/cohd/openshift-rmlstreamer-cohd-associations.nt
nohup cat * >> openshift-rmlstreamer-cohd-associations.nt &

ls -alh /mnt/cohd/openshift-rmlstreamer-cohd-associations.nt/openshift-rmlstreamer-cohd-associations.nt

# Zip the merged output file:
nohup gzip openshift-rmlstreamer-cohd-associations.nt &

Copy to your server#

SSH connect to your server, http_proxy var might need to be changed temporarily to access the DSRI

export http_proxy=""
export https_proxy=""

# Copy with oc tool:
oc login
oc cp flink-jobmanager-7459cc58f7-cjcqf:/mnt/cohd/openshift-rmlstreamer-cohd-associations.nt/openshift-rmlstreamer-cohd-associations.nt.gz /data/graphdb/import/umids-download &!

# Check (19G total):
ls -alh /data/graphdb/import/umids-download
cp /data/graphdb/import/umids-download/openshift-rmlstreamer-cohd-associations.nt.gz /data/d2s-project-trek/workspace/dumps/rdf/cohd/
gzip -d openshift-rmlstreamer-cohd-associations.nt.gz

Reactivate the proxy (EXPORT http_proxy)

Preload in GraphDB#

Check the generated COHD file on the server at:

cd /data/d2s-project-trek/workspace/dumps/rdf/cohd

Replace wrong triples:

sed -i 's/"-inf"^^<http:\/\/www.w3.org\/2001\/XMLSchema#double>/"-inf"/g' openshift-rmlstreamer-cohd-associations.nt

Start preload:

cd /data/deploy-ids-services/graphdb/preload-cohd
docker-compose up -d

The COHD repository will be created in /data/graphdb-preload/data, copy it to the main GraphDB:

mv /data/graphdb-preload/data/repositories/cohd /data/graphdb/data/repositories