Building and Using Large Common Sense Knowledge Bases

A project in the DARPA High Performance Knowledge Base Initiative

Principal Investigator: Kenneth D. Forbus

Research Scientist: Ron Ferguson

We believe that effectively creating and using large-scale domain theories (10,000 - 100,000 axiom equivalents) will require improving the bandwidth of human-computer interaction in several dimensions: the kinds of concepts used, the reasoning methods, and the interaction media. More specifically,

• At the conceptual level, qualitative representations provide an ideal medium for expressing the framework knowledge of the physical world that both provides simple inferences directly and also organizes and orchestrates the use of more detailed knowledge.
  
  
• At the level of reasoning methods, analogical reasoning techniques based on the best available theories in cognitive science can provide more human-like case-based reasoning systems.
  
  
• At the level of interaction media, the use of sketching can provide a more natural way for domain experts and software to communicate about the spatial aspects of concepts and situations.
  
  

These levels strongly interact. By tackling the right aspects of all three at once, as part of a community that is generating large-scale domain theories, we believe we can generate breakthrough capabilities. By incorporating these capabilities into a development environment, we can transfer this technology to other developers and users. Specifically,

• We propose to create a set of fundamental domain theories, focusing on common sense physical phenomena, that can be used as components in knowledge bases for multiple purposes, including effects-based planning and intelligence analysis tasks. These fundamental domain theories will include aspects of space, quantity, time-varying behavior, substances, physical processes, and causality. We will use these domain theories to develop higher-level domain theories, include developing the common sense concepts needed to model phenomena involving weather and climate, civilian logistics distribution systems, and basic ideas of vehicles.
  
  
• We propose to develop a set of analogical processing systems that can serve as a foundation for a new generation of case-based reasoning systems. Our goals are to provide retrieval times in the 10-20 second range even on 10,000 entry libraries of structured propositional representations, and to enable domain experts to extend CBR systems without having to understand the system's internals.
  
  
• We propose to develop sketching as a tool for knowledge acquisition and refinement. Our goal is to make sketching as natural a modality for explaining things to machines as it is for explaining things to people, concentrating on the specific tasks of defining new spatial recognizers and interactive explanation of concepts with spatial aspects.
  
  
• We propose to create a Domain Theory Development Environment. Our goal is to create an environment that can handle 100,000 axiom-equivalent knowledge bases in a high-end networked computational environments, but is still fieldable on notebook computers for medium-scale tasks and knowledge bases, allowing the environment to be brought to the domain experts as well as the other way around.

---

Back to Symbolic Supercomputing page | Back to Structure-Mapping: A computational model of analogy and similarity page | Back to Sketching: Towards a computational model page | Back to Qualitative Process Theory page | Back to MAC/FAC: A model of similarity-based retrieval page | Back to Compositional Modeling page | Back to Previous Projects page | Back to QRG Home Page