Research Project Name

Learning by Natural Instruction: Apprenticeship Learning, Critiquing, and Question-Answering


The overall objective of this project is a cognitively-plausible computational model of knowledge-intensive apprenticeship learning expertise in medical decision-making domains. This objective is being achieved by experimenting with different knowledge representation and inference architectures, and measuring the extent that they naturally support the "multiple dimensions of expertise", such as apprenticeship learning, critiquing, and question-answering.


How does one capture the architecture of human higher-level cognition for knowledge-intensive decision making? What is the correct level of abstraction? Our approach originated from the observation that when an expert is able to solve a decision-making problem, they can immediately use their expertise in a large variety of  ways.

We refer to the large number of ways that an expert can use their expertise as the "multiple dimensions of expertise". The ones we have concentrated on are as follows. First, problem solving, such as medical diagnosis and ship damage control. Second, teaching their expertise, by presenting problems to students and critiquing the students solutions. Third, critiquing other experts who are solving a problem in the same domain. Fourth, apprenticeship learning, wherein  they observe problem solving or their problem solving is observed, such as a medical internship. Fifth, a post-session question-answering
spoken dialogue, wherein the ask or answer questions in a dialogue with another expert. Sixth, learning by resolving  gaps in its expertise during problem solving, by using induction over past problem cases they have seen. These "multiple dimensions of expertise" are usually free side-effects of learning to solve problems at an expert level. They goal is to produce an integrated architecture that yields them as a free side-effect.


Over the last 25 years, a succession of systems have been implemented that progressively have come closer to the overall objective. They include Odysseus, Odysseus2, Minerva, TIPN, and Gerona. The major challenge has been to find a method of knowledge representation and inference that yields the desired multiple dimensions of expertise. The key issue has been learnability (by apprenticeship, critiquing, Q/A) via natural instruction.

Selected Publications