Machine learning from casual conversation

• Published in: Machine learning Vol. 112; no. 12; pp. 4789 - 4836
• Main Authors: Mohammed Ali, Awrad E., Gonzalez, Avelino J.
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.12.2023; Springer Nature B.V
• Subjects: Artificial Intelligence; Cloning; Computer Science; Control; Intelligent agents; Knowledge bases (artificial intelligence); Machine Learning; Mechatronics; Natural language; Natural language processing; Natural Language Processing (NLP); Robotics; Simulation and Modeling; Verbal communication; Written language; Chatbot; Natural language processing; Learning from conversation; Memory update; Machine learning; Classification
• ISSN: 0885-6125; 1573-0565
• DOI: 10.1007/s10994-023-06383-0

Human social learning is an effective process that has inspired many existing machine learning approaches, such as learning from observation and learning by demonstration . In this paper, we introduce another form of social learning, learning from a casual conversation or LCC a machine learning approach in which an artificially intelligent agent learns new information through an extended natural language dialog with a human. Our system enables the agent to add or change information in its knowledge base as a result of the human’s conversational text inputs. LCC seeks to close an important gap in the state of the art that has focused on teaching computer agents how to perform specific tasks. Furthermore, LCC could also provide an efficient way to enhance the knowledge base of certain types of systems without requiring the involvement of a programmer. LCC does not require the user to enter specific information; instead, the user can converse naturally with the agent. As part of its learning process, LCC identifies the text inputs from the conversing human that contain information worth learning, and then determines whether the inputs are heretofore unknown and learns it; in agreement with what it already “knows” and ignores it; or in conflict with what it “knows” and it must resolve the conflict. LCC’s architecture consists of multiple sub-systems combined to perform the above tasks. Its learning component can add new information to the knowledge base, confirm existing information, and/or update existing information found to be related to the user input. The LCC system functionality was rigorously assessed with test statements comprising various difficulty levels. Furthermore, its acceptance by human users was evaluated by two separate groups of human test subjects—one group who interacted with the system, and a second group that evaluated the logs of the interactions of the first group. The collected results were all found to be acceptable and within the range of our expectations.