Awesome
Knowledge Graph Construction with R2RML and RML
We conduct an evaluation of KGC engines considering several R2RML and RML processors to identify their strengths and weaknesses. We (i) perform a qualitative analysis of the distinctive features of each engine, (ii) examine their conformance with the mapping language specification they support, and (iii) assess their performance and scalability using the GTFS-Madrid-Bench benchmark.
Citing:
@inproceedings{arenas2021knowledge,
title = {{Knowledge Graph Construction with R2RML and RML: An ETL System-based Overview}},
author = {Arenas-Guerrero, Julián and Scrocca, Mario and Iglesias-Molina, Ana and Toledo, Jhon and Pozo-Gilo, Luis and Doña, Daniel and Corcho, Oscar and Chaves-Fraga, David},
booktitle = {Proceedings of the 2nd International Workshop on Knowledge Graph Construction},
year = {2021},
series = {CEUR Workshop Proceedings},
publisher = {CEUR-WS.org},
volume = {2873},
url = {http://ceur-ws.org/Vol-2873/paper11.pdf},
}
Engines
We test the performance and scalability of a set of KG construction engines:
R2RML-based:
- Ontop v4.1.0
- Morph-RDB v3.12.5
- R2RML-F v1.2.3
- db2triples v2.2
RML-based:
- RMLMapper v4.9.1
- CARML v0.3.2
- RocketRML v1.8.2
- SDM-RDFizer v3.5
- RMLStreamer v2.0
- Chimera v2.1
Evaluation resources
GTFS-Madrid-Bench
Using the GTFS-Madrid-Bench and based on the input dataset we create the following distributions to test the engines:
- Formats: CSV, XML, JSON, RDB and Random-Custom (sources in different formats)
- Scale-Sizes: 1, 10, 100 and 1000
Data can be directly download executing bash scripts/download-data.sh
R2RML and RML test-cases
We use the resources provided by the W3C community on KG-Construction (https://www.w3.org/community/kg-construct/) to run the R2RML and RML test-cases over the selected engines.
Results
We created a comparative framework to gather and compare the information about the engines' features, availabe here; and tested the engines with the mentioned benchmark in terms of time and memory used. The raw data resulting from the evaluation is stored here, and the resulting figures can be seen here.
Authors
- Julián Arenas-Guerrero - julian.arenas.guerrero@upm.es (Ontology Engineering Group - UPM)
- Mario Scrocca (Cefriel - Politecnico di Milano)
- David Chaves-Fraga (Ontology Engineering Group - UPM)
- Jhon Toledo (Ontology Engineering Group - UPM)
- Daniel Doña (Ontology Engineering Group - UPM)
- Luis Pozo-Gilo (Ontology Engineering Group - UPM)
- Ana Iglesias (Ontology Engineering Group - UPM)