Short Answer Questions Generation by Fine-Tuning BERT and GPT-2

Danny C.L. Tsai, Willy J.W. Chang, Stephen J.H. Yang

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

In educational research, artificial intelligence (AI) is suitable for many situations, such as exploring student learning paths and strategies. However, most of them cannot reduce the workload of teachers. In the course, teachers need to spend a lot of effort on setting exams, because exams are the most direct way to understand students' learning performance. In this research, we use modern artificial intelligence model, BERT and GPT-2 to generate questions to reduce the work of teachers frequently setting questions. The type of questions we generate is short answer questions. The main reason is that many researches prove that short-answer questions can enhance students' long-term memory and improve learning performance. We also compare the performance of BERT before and after fine-tuning. The results show that BERT can be used for general reading comprehension questions before fine-tuning, but in the field of domain knowledge, fine-tune BERT's performance is better.

Original languageEnglish
Title of host publication29th International Conference on Computers in Education Conference, ICCE 2021 - Proceedings
EditorsMaria Mercedes T. Rodrigo, Sridhar Iyer, Antonija Mitrovic, Hercy N. H. Cheng, Dan Kohen-Vacs, Camillia Matuk, Agnieszka Palalas, Ramkumar Rajenran, Kazuhisa Seta, Jingyun Wang
PublisherAsia-Pacific Society for Computers in Education
Pages508-514
Number of pages7
ISBN (Electronic)9789869721486
StatePublished - 22 Nov 2021
Event29th International Conference on Computers in Education Conference, ICCE 2021 - Virtual, Online
Duration: 22 Nov 202126 Nov 2021

Publication series

Name29th International Conference on Computers in Education Conference, ICCE 2021 - Proceedings
Volume2

Conference

Conference29th International Conference on Computers in Education Conference, ICCE 2021
CityVirtual, Online
Period22/11/2126/11/21

Keywords

  • Automatic question generation
  • BERT
  • Keyword extraction
  • Learning performance

Fingerprint

Dive into the research topics of 'Short Answer Questions Generation by Fine-Tuning BERT and GPT-2'. Together they form a unique fingerprint.

Cite this