Knowledge bases for explainable benchmarking (QALD10, QALD9+DB, QALD9+WK)

This project provides three knowledge graphs that we created for the three QA benchmarks: QALD-9 plus DBpedia, QALD-9 plus Wikidata, and QALD-10. Here are some more details:

1. Preprocessing

• We remove all questions from the three QA datasets that have an empty ground truth answer set.

• We preprocessed the DBpedia reference graph by:

  • Removing 43,618 triples with IRIs that do not pass through the RDF checker.

  • Removing properties of the http://dbpedia.org/property/ namespace.

  • Inferring the classes of all entities based on the class hierarchy.

• We preprocessed Wikidata by replacing the property http://www.wikidata.org/prop/direct/P31 with http://www.w3.org/1999/02/22-rdf\textbackslash-syntax-ns\#type.

2. Knowledge Base Structure

In the first step of our benchmarking framework, we generate a knowledge graph comprising information from the dataset used during the benchmarking process. Our work relies on the QALD datasets, which include three types of data for each question:

1. Natural language question
   Each question comes with a representation in several languages. From the English question, we extract linguistic features such as:

   • The length of the question (dqb:hasLength)                                                                                                                      Note: The prefix dqb: refers to the namespacehttp://w3id.org/dice-research/qa-bench#.

   • The presence of negation (dqb:hasNegation)

   • The question word (dqb:hasQuestionWord)

   • The NLP parse tree (dqb:hasNlpParseTreeRoot)
     Note: We employ the Stanford NLP toolkit for the extraction.

2. Answer(s)
   Each question comes with the ground truth answers. We add these answers to the generated graph with three different properties distinguishing:

   • IRI answers (dqb:hasIRIAnswer)

   • Boolean answers (dqb:hasBooleanAnswer)

   • Other literal answers (dqb:hasLiteralAnswer)
     For each IRI listed as an answer, we add its concise bounded description (CBD) extracted from the reference knowledge graph.

3. SPARQL query
   Each question has a SPARQL query that returns the ground truth answer when used on the reference knowledge graph. We adopt LSQ to add the following SPARQL query features to our knowledge graph:

   • Entities (dqb:hasEntity), properties (dqb:hasProperty) contained in the query, and the CBD of the entities

   • Type of query

   • The number of triple patterns

   • The number of basic graph patterns

   • The average degree of vertices

   • The median degree of vertices involved in join operations

   • The minimum, maximum, and median number of triple patterns in a basic graph pattern

   • The presence of certain keywords such as FILTER, DISTINCT, and GROUP BY