A new approach to generate teacher-like questions guided by text spans extraction
Résumé
Generating teacher-like questions and answers remains an open issue while being useful for students, teachers and teaching aid application providers. Given a textual course material, we are interested in generating non-factual questions which require an elaborate answer (implying some sort of analysis or reasoning). Despite the availability of annotated corpora of questions and answers, two main obstacles prevent the development of such generator using deep learning. Firstly, the amount of qualitative data is not sufficient to train generative approaches. Secondly, for a stand-alone application, we do not have an explicit support to guide the generation towards complex questions. In this article, we propose and compare several new retargetable language algorithms for answer text span support extraction and complex question generation, on secondary education course material use-case in French. We study the contribution of deep neural syntactic parsing and transformer based semantic representation, relying on the question type (according to our specific question typology) and the support text span in the context. We highlight the important role of nominal noun phrases and dependency relations, as well as the gain brought by recent transformer language models.
Domaines
Informatique [cs]Origine | Fichiers produits par l'(les) auteur(s) |
---|