Modeling question asking using neural program generation
Project Overview
The document explores the application of a neural program generation framework designed to emulate human question-asking behavior in an educational context, particularly through an information-search game akin to 'Battleship'. By integrating deep neural networks with symbolic programming, this innovative framework generates creative and informative questions, significantly surpassing traditional question-generation methods. It can learn from human demonstrations or operate autonomously via reinforcement learning, showcasing its adaptability and effectiveness. The findings reveal that the model successfully synthesizes questions that mirror human inquiry, highlighting its potential to enhance educational experiences by fostering critical thinking and inquiry-based learning. This advancement signifies a promising direction for the use of generative AI in education, aiming to improve student engagement and learning outcomes through the art of questioning.
Key Applications
Neural program generation framework for question asking
Context: Educational context involving creative question generation in information search games, targeting learners engaging in inquiry-based learning.
Implementation: The model is trained using a combination of supervised learning with human demonstrations and reinforcement learning without human examples, leveraging a convolutional encoder and transformer decoder.
Outcomes: The framework successfully generates human-like questions and adapts to various contexts, showing improved performance over previous methods.
Challenges: The model has limitations in generalizing to different scenarios and occasionally produces meaningless questions.
Implementation Barriers
Technical barrier
The model cannot generalize to systematically different scenarios than it was trained on. Additionally, the model sometimes generates meaningless questions.
Proposed Solutions: Future work aims to explore the model’s compositional abilities and joint training for question asking and answering. Enhancements to the grammar-based framework and further training methodologies are proposed to improve output quality.
Project Team
Ziyun Wang
Researcher
Brenden M. Lake
Researcher
Contact Information
For information about the paper, please contact the authors.
Authors: Ziyun Wang, Brenden M. Lake
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