RML (https://rml.io/specs/rml/) is a mapping language to map data from various data sources (CSV, JSON, RDBMS, …) to RDF. It extends R2RML (https://www.w3.org/TR/r2rml/) which only supports RDBMS sources.

This project is about:

• Developing a tool to automatically create RML files given a data source
• Evaluating the quality of the produced RML files on downstream tasks, such as question answering, knowledge graph entity linking, etc.

This project is not about:

• Creating a tool to convert data given a RML file. For this, there already exist a number of popular tools, such as
  • https://github.com/morph-kgc/morph-kgc (for materialized RDF from various data sources)
  • https://github.com/ontop/ontop (for a virtual knowledge graph setup over relational databases)
  • https://github.com/tarql/tarql (not via RML, but another way to map from CSV to RDF)
  • https://github.com/Swirrl/csv2rdf (also not via RML, but also CSV to RDF)

Challenges:

• Coming up with a semantically meaningful schema for the RDF data is difficult; possible approaches could be
  • use LLMs or other world models to construct the RML
  • combine the above with a human-in-the-loop approach
  • combine the above with a RL approach
• We do not want to just convert the input data source trivially to RDF, e.g. for a CSV file one could
  • convert each row to <row_i> <type> <Row>
  • and each column in that row to <row_i> <col_name> <col_value>

Relational data:

Persons

Movies

Actor

Bad RDF: