Knowledge Acquisition via Analogy, Examples, and Sketching
Principal Investigator: Kenneth D. Forbus
Research Scientist: Ron Ferguson
The goal of this project is to use ideas from cognitive science to develop a set of ideas, techniques, and components that enable the construction of systems that domain experts can use to interactively create, inspect, extend, and maintain large scale knowledge bases. Cognitive science is making solid progress in discovering how people reason with and learn from examples, how people use analogy and metaphor to communicate, and how people use their visual systems to help them understand the world. Using techniques from AI and other areas of computer science, we have been using these ideas to create technologies that come closer in performance to human-style breadth and flexibility. In this project we will apply this approach to the problems of Rapid Knowledge Formation in three areas:
- Reasoning and learning from examples and analogies. We will
develop an analogical abduction system
that enables expert-supplied examples to be retrieved and used to solve new
problems, without the hand-tailoring of indexes or the domain-specific
algorithms required by today's case-based reasoning technology. We will develop an Examples Memory System that combines conservative, incremental
learning with analogical abduction, to acquire generalizations from multiple
- Using analogy and metaphor
in communication. We will develop techniques for interpreting
sentence-level metaphors, to enable experts to communicate more naturally
with knowledge acquisition systems. We will develop techniques for interpreting
explanatory analogies so that experts can engage in extended analogies in
dialogs with knowledge acquisition systems.
- Sketching for knowledge acquisition. Sketching is a natural way people use to explain and apprehend complex ideas, especially when they involve space (literally or metaphorically). Our goal is to push the state of the art to the point where experts and knowledge acquisition systems can interact via sketching. This requires developing a spatial reasoning engine that provides human-like perceptual services, a graphical symbology domain theory that describes how visual and spatial information are used to convey meaning, an everyday physical semantics domain theory that specifies aspects of common sense physics commonly used in sketches, and sketch generation and modification techniques that will enable knowledge acquisition software to graphically express its own knowledge when interacting with experts.
These ideas, techniques, and components will be tested in two ways. First, they will be used in end-to-end systems being created as integrated team efforts under RKF, through subcontracts with two integrated teams (Cycorp and SRI). Second, we will create a sketching Knowledge Entry Associate (sKEA) that uses knowledge-rich multimodal interface techniques to interact with experts via sketching. sKEA will also be provided to collaborating integrated teams, to provide a complement to the forms-based and text/hypertext-based interfaces they are creating.