Generating Answerable Questions from Ontologies

Proposals for automating the creation of teaching materials across the sciences and humanities include question generation from ontologies.

Those efforts have focused on multiple-choice questions, whereas learners also need to be exposed to other types of questions, such as yes/no and short answer questions. Initial results showed it is possible to create ontology-based questions. It is unknown how that can be done automatically and whether it would work beyond that use case in biology.

We investigated this for ten types of educationally useful questions with additional sentence formulation variants. Each type of questions has a set of template specifications, axiom prerequisites on the ontology, and an algorithm to generate the questions from the ontology.

Three approaches were designed: template variables using foundational ontology categories, using main classes from the domain ontology, and sentences mostly driven by natural language generation techniques.

The user evaluation showed that the second approach resulted in slightly better quality questions than the first, and the linguistic-driven templates far outperformed both on syntactic and semantic adequacy of the questions.