Skip to content

DBpedia's GSoC warmup tasks

Tommaso Soru edited this page May 5, 2020 · 2 revisions

Hi! If you landed on this page, you are likely interested in becoming a GSoC student for DBpedia. In the following, the warmup tasks are described in details.

Warmup tasks for A Neural QA Model for DBpedia

  • Read the papers:
  • Train a Neural SPARQL Machine model on a DBpedia class of your choice.
    • In this example, we are going to select dbo:Monument as the class of instances to perform QA on.
    • In the template file data/annotations_monument.csv, you may find a collection of query templates. The columns show the following values:
      • Column 1: target class of the first placeholder <A> in a question (e.g., dbo:Monument)
      • Column 2: empty for special use
      • Column 3: target class of the second placeholder <B> in a question, if it exists (e.g., dbo:Monument)
      • Column 4: question template in natural language (e.g., is <A> more recent than <B>)
      • Column 5: SPARQL query template which translates column 4 (e.g., ask where { <A> dbp:complete ?a . <B> dbp:complete ?b . FILTER(?a > ?b) })
      • Column 6: SPARQL generator query to collect examples for placeholders <A>, <B>, etc. (e.g., select distinct(?x) ?y where { ?x dbp:complete [] . ?y dbp:complete [] . FILTER(?x != ?y) })
    • Create a template file for a class of your choice with at least 20 query templates. Hints:
      • Mind the correct spacing in the SPARQL query.
      • Use all-lowercase letters in questions except for placeholders.
      • To speed up experiments, choose one having less than 1,000 instances.
    • Follow the data preparation, training, and inference sections in the readme file.
  • Share your results! We are curious how well NSpM perform on each use-case.
Clone this wiki locally