Jill Watson: A Virtual Teaching Assistant powered by ChatGPT
Project Overview
The document discusses the integration of generative AI in education, highlighting the development of Jill Watson, a Virtual Teaching Assistant (VTA) at Georgia Institute of Technology, which utilizes ChatGPT to provide real-time, personalized support to students. By effectively addressing inquiries related to course content and logistics, Jill Watson employs a sophisticated architecture that combines various AI techniques to improve response accuracy and mitigate issues like hallucination and toxicity. The system has shown significant advancements compared to previous AI iterations and traditional educational tools, resulting in enhanced student engagement and improved learning outcomes. Overall, the findings suggest that generative AI applications like Jill Watson can transform educational experiences by offering continuous support, fostering a more interactive learning environment, and ultimately benefiting both students and educators.
Key Applications
Jill Watson, a Virtual Teaching Assistant powered by ChatGPT
Context: Online classrooms at Georgia Institute of Technology, targeting students enrolled in various courses
Implementation: Utilizes a modular design to integrate ChatGPT with a skill-based architecture, employing dense passage retrieval (DPR) for answering queries based on course documents.
Outcomes: Increased reliability in responding to student queries, reduced instances of harmful or confusing answers, and improved engagement through conversational interaction.
Challenges: Limited to the context of provided documents, reliance on the accuracy of each module, and potential for the skill classifier to misidentify relevant queries.
Implementation Barriers
Technical/Safety barrier
Challenges related to the accuracy and reliability of the AI in understanding and responding to complex student queries, alongside the risk of generating inappropriate or toxic responses that can harm students.
Proposed Solutions: Employing safety measures such as moderation filters and classifier systems to manage response quality and prevent harmful content in responses.
Performance barrier
The effectiveness of the system relies on the proper functioning of multiple modules, where failures in one can impact the overall performance.
Proposed Solutions: Designing modular systems to allow for easy updates and improvements to individual components.
Project Team
Karan Taneja
Researcher
Pratyusha Maiti
Researcher
Sandeep Kakar
Researcher
Pranav Guruprasad
Researcher
Sanjeev Rao
Researcher
Ashok K. Goel
Researcher
Contact Information
For information about the paper, please contact the authors.
Authors: Karan Taneja, Pratyusha Maiti, Sandeep Kakar, Pranav Guruprasad, Sanjeev Rao, Ashok K. Goel
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