Title:	Multiple-choice question generation in Swedish
DNr:	Berzelius-2022-169
Project Type:	LiU Berzelius
Principal Investigator:	Johan Boye <jboye@kth.se>
Affiliation:	Kungliga Tekniska högskolan
Duration:	2022-08-31 – 2022-12-01
Classification:	10208
Homepage:	https://www.csc.kth.se/~jboye/swequest.html
Keywords:

Abstract

Multiple-choice questions (MCQs) provide a way of assessing reading comprehension on many educational levels. Such an MCQ consists of a text, a question related to the context, and a set of answer alternatives. One of these alternatives is the correct answer, whereas the others are “distractors” -- incorrect but plausible answers to the stem. In this project, we will train a Transformer-based model for generating such MCQs automatically from a given Swedish text, using a large Swedish text corpus and a smaller, specialized corpus of MCQs which we have collecter. Being able to automatically generate such MCQs from a given Swedish text would entail many advantages for teachers, who can spend less time on text construction and more on teaching, and for students who could use such a system for self-studies. An important aspect of any such system is that the correct answer is indeed correct given the text and distractors are indeed wrong (although plausible). Achieving both of these goals requires a lot of experimentation on our part: developing, tuning and comparing different neural models for the generation of texts and questions with associated answers and distractors.