article Article Summary
Nov 07, 2024
Blog Image

An automatic test question generation system combining natural language processing and deep learning techniques successfully generates high-quality educational questions, outperforming existing models in both automated and human evaluations.

Objective: To develop and evaluate an automatic question generation system that can create high-quality examination questions from teaching materials using natural language processing and Topic Model, focusing on selecting meaningful declarative sentences and generating relevant questions.

Methods:

  • Two-stage approach combining sentence selection and neural question generation
  • Implementation of topic-embedding question generation (TE-QG) model
  • Use of multisource teaching materials for sentence selection
  • Experimental evaluation using SQuAD dataset
  • Both objective metrics (BLEU, ROUGE-L) and subjective assessment (human evaluation)
  • Comparison with baseline models including H&S, L2A, and NQG

Key Findings:

  • The proposed Centroid_Emb method outperformed existing sentence selection methods
  • TE-QG model achieved better performance than baseline models across all metrics
  • Topic feature integration improved question generation quality
  • Filtering parts of speech in topic words enhanced model performance
  • Human evaluators rated TE-QG-generated questions higher in fluency, clarity, and usefulness

Implications: The system can help reduce teachers' workload in creating exam questions while providing students with high-quality practice materials for self-learning, contributing to more efficient and personalized educational assessment.

Limitations:

  • Limited to single version/publisher of teaching materials
  • Focuses only on textual content, excluding images and tables
  • Restricted to factual questions
  • Dependent on quality of input teaching materials

Future Directions:

  • Explore innovative algorithms to improve efficiency
  • Incorporate multiple versions/publishers of teaching materials
  • Include image and table analysis
  • Consider external resources for improved term comparison
  • Develop methods to handle synonymous and proprietary terms

Title and Authors: "A Method for Generating Course Test Questions Based on Natural Language Processing and Deep Learning" by Hei-Chia Wang, Yu-Hung Chiang, and I-Fan Chen

Published on: 2024

Published by: Education and Information Technologies

Related Link

Comments

Please log in to leave a comment.