Learning from natural tutorial instruction


The aim of this project is to develop an electronic student that can learn from instruction that is increasingly more natural (i.e. instruction that increasingly resembles the instruction that a human teacher provides a human student).



We are currently focusing on two aspects of this work:
  1. Learning Procedures from Tutorial Instruction, where we extend the Wings Workflow System to enable users to teach new procedures using natural language instructions. We have run formative user studies.
  2. A Framework for Combining Instruction and Demonstration, where we investigate the formal requirements and possibilities for combining these two instruction methods to teach more effectively. We have implemented a prototype of the formal framework we have developed and tested it on a number of example instructions together with demonstrations for teaching new game playing procedures on a real-time strategy game.



Natural instruction is inherently error prone and incomplete. Human teachers regularly omit important details, either because for a human such details might be obvious, or because they simply forgot about them. We are developing a system that is able to semantically reason about the consistency of instruction, capable of hypothesizing ways to correct errors or fill gaps in instructions, and rank such hypotheses in terms of their likelihood.

More -->


More publications -->


See a demo -->


Group Members: Alumni and Former Members:

More -->


  • DARPA / Bootstrapped Learning

More -->


More links -->

<< Back to IKCAP