Pre-Training With Scientific Text Improves Educational Question Generation
Project Overview
The document explores the integration of generative AI in education, focusing on the development of EduQG, an innovative educational question generation model that utilizes a large language model to enhance the automatic creation of educational questions. This advancement is significant for promoting self-assessment in personalized learning environments, allowing learners to engage more effectively with their studies. The findings from initial experiments suggest that pre-training EduQG on scientific texts significantly improves its capability to generate high-quality educational questions, thereby supporting educators and learners alike. Overall, the implementation of generative AI technologies like EduQG represents a promising approach to enriching educational practices, fostering individualized learning experiences, and enabling learners to assess their understanding more effectively.
Key Applications
EduQG - a novel educational question generation model
Context: Educational context focused on self-assessment and personalized learning; target audience includes learners utilizing digital educational resources.
Implementation: The model was implemented by fine-tuning a large language model (T5) with scientific text datasets (S2ORC and SciQ) to enhance question generation capabilities.
Outcomes: EduQG showed improved predictive performance in generating educational questions compared to the baseline model (Leaf), as indicated by better BLEU and F1 scores.
Challenges: The linguistic quality of generated questions did not meet expectations, possibly due to a mismatch in language style and vocabulary between scientific texts and the reference models.
Implementation Barriers
Technical Barrier
Mismatch between the complex vocabulary of scientific language and the language models used for linguistic quality assessment.
Proposed Solutions: Future work will address the linguistic quality issues through deeper analysis and potentially using both offline and human evaluations.
Project Team
Hamze Muse
Researcher
Sahan Bulathwela
Researcher
Emine Yilmaz
Researcher
Contact Information
For information about the paper, please contact the authors.
Authors: Hamze Muse, Sahan Bulathwela, Emine Yilmaz
Source Publication: View Original PaperLink opens in a new window
Project Contact: Dr. Jianhua Yang
LLM Model Version: gpt-4o-mini-2024-07-18
Analysis Provider: Openai