Evorus: A Crowd-powered Conversational Assistant Built to Automate Itself Over Time
Project Overview
The document discusses the innovative application of generative AI in education through Evorus, a crowd-powered conversational assistant that evolves by incorporating crowd-generated responses alongside automated components. This hybrid approach leverages human oversight to enhance the quality of interactions while progressively decreasing dependence on crowd input. The architecture of Evorus facilitates seamless integration of new chatbots and automated voting mechanisms, which significantly boosts efficiency and cost-effectiveness in real-time educational conversations. Key applications of this technology include personalized tutoring, interactive learning environments, and enhanced student engagement. Findings indicate that such systems can improve learning outcomes by providing tailored support and immediate feedback to learners. The combination of human and AI elements not only enhances the conversational experience but also allows for continuous improvement and adaptation of the assistant's capabilities, making it a transformative tool in the educational landscape. Overall, the document underscores the potential of generative AI to revolutionize educational practices by fostering more dynamic and responsive learning experiences.
Key Applications
Evorus, a crowd-powered conversational assistant
Context: Open-domain dialog system for users interacting via Google Hangouts
Implementation: Deployed for 5 months with participant feedback to refine the system
Outcomes: Automated response usage increased, costs reduced by 32.76%, and user satisfaction ratings remained high.
Challenges: Balancing automation and crowd participation; ensuring response quality.
Implementation Barriers
Operational Challenge
High monetary costs and response latency associated with crowd-powered systems.
Proposed Solutions: Gradual automation through a hybrid system that integrates both crowd and machine components.
Quality Control
Potential for low-quality automated responses affecting conversation quality.
Proposed Solutions: Crowd oversight to verify responses and machine learning to improve response selection.
Project Team
["Ting-Hao "Kenneth" Huang", "Joseph Chee Chang", "Jeffrey P. Bigham"]
Researcher
Contact Information
For information about the paper, please contact the authors.
Authors: ["Ting-Hao "Kenneth" Huang", "Joseph Chee Chang", "Jeffrey P. Bigham"]
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