Towards Teachable Conversational Agents
Project Overview
The document examines the application of generative AI in education, specifically through the development of teachable conversational agents capable of learning from human feedback during interactions. It highlights the promise of interactive machine learning systems to enhance educational experiences, particularly in improving performance in tasks such as text classification. The findings suggest that involving users in the teaching process can significantly boost the agents' classification capabilities, although it also points out existing challenges in identifying effective teaching strategies and the influence of varied teaching inputs. Overall, the document underscores the potential of conversational interfaces to foster better learning outcomes while recognizing the complexities involved in their implementation.
Key Applications
Teachable conversational agent for text classification
Context: Web-based learning environment for teaching classification of news articles
Implementation: Participants engage in teaching the conversational agent by conversing about articles, asking questions, and providing feedback.
Outcomes: Improved performance of the agent in classifying articles based on human feedback; performance increased with the number of dialogues exchanged.
Challenges: Performance can degrade when external words are taught; effectiveness varies based on the quality of the teacher's input.
Implementation Barriers
Technical barrier
Quality of word embeddings used for the conversational agent is limited due to using a smaller dataset.
Proposed Solutions: Future work could focus on using contextual embeddings like BERT trained on more relevant datasets.
Human factor barrier
Not all human teachers provide effective input, leading to variable performance of the agent.
Proposed Solutions: Investigate characteristics of effective teachers and factors influencing teaching quality.
Project Team
Nalin Chhibber
Researcher
Edith Law
Researcher
Contact Information
For information about the paper, please contact the authors.
Authors: Nalin Chhibber, Edith Law
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