Additional Slide Presentations (no/multiple associated papers)

The following are additional slide presentations/discussions I've given, which either don't have an associated paper, or are associated with multiple papers:

• Building Large Knowledge Bases from Reusable Components (Peter Clark, 1998). This presentation is a good summary of the work I have been doing with Bruce Porter on KB architectures, described in various papers.

  Abstract:
  As the world of structured information grows (eg. databases, KBE, XML, electronic commerce, ...) there are tremendous opportunities for applying knowledge-based systems technology (eg. search, question-answering, information retrieval). However to achieve robust, multi-purpose behaviour, such systems require a vast amount of world knowledge -- and building and maintaining large-scale knowledge-bases (KBs) is a notoriously difficult task.

  In response to this, our goal is to instead be able to construct large-scale, knowledge-based systems through the assembly of prefabricated, representational components. Each component encapsulates a small, reusable theory about the world, describing a recurring abstraction (eg. production, distribution, mechanical device); domain-specific representations are then constructed automatically by composing such abstractions together. This approach provides simplicity and modularity in the KB; in addition the system can compose representations of and reason with novel, unanticipated concepts at run-time. In this talk, I will describe a KB architecture based on this approach, some component theories, the composition mechanisms for integrating them, and a simple prototype built recently using the library. Finally I'll speculate on possible, longer-term developments of this work.

• Some Desiderata for Machine Understanding (Peter Clark, May 2002, Talk at Univ Texas at Austin)

  Abstract:
  In this informal "brainstorming" talk, I explore five general requirements needed for machine understanding of fragmented, noisy input (in particular, but not only, natural language: Elaboration ("scene building"), representing possibilities, coherence assessment, viewpoints/context, and knowledge acquisition.

• The Myth of Commonsense Knowledge (Peter Clark, 2004, for Boeing's KR discussion group)

  Abstract:
  The history of science shows that, at times, an entire community can make assumptions which are so fundamental, and so universally shared, that they become almost subconsious, and thus rarely get questioned or even noticed. One such assumption in the field of AI is what I will call "the myth of common sense", namely the belief that the key to building intelligent systems lies in the accumulation of, or access to, a vast body of world knowledge. Some groups, such as Cycorp, have pursued this objective by attempting to hand-craft such a repository; more recently, others are exploring the creation of such such repositories semi-automatically, e.g., by applying statistics/text mining/info retrieval techniques to large corpora of text, or by community-based knowledge accumulation; finally, a sizeable proportion of the field believe systems ultimately need lots of knowledge, but consider this too difficult to achieve, and so spend their time working on something else instead. In all these cases, however, there is a shared assumption that lots of knowledge is the solution, and that building it is the problem. The "knowledge acquisition bottleneck", first identified as a grand challenge in AI over 30 years ago, is still almost universally accepted as the fundamental barrier to machine intelligence. In this mixed presentation/discussion, I'll review how ubiquitous this assumption is, challenge it, and suggest some alternative issues which are perhaps more fundamental for progress.

• The Knowledge Acquisition Bottleneck Revisited: How can we build large KBs? (Peter Clark and John Thomson, 2004, for Boeing's KR discussion group)

  Abstract:
  How can a computer acquire the massive amount of knowledge needed for "intelligent" reasoning? There are a variety of techniques people are exploring**, but it's not clear which are good, which are bad, and which are just plain ugly. I'll briefly summarize some of these approaches, and then facilitate some discussion of them to explore which ones people see hope in, and which seem to be heading for a dead end.

  [** e.g., hand-code (Cyc, WordNet, FrameNet), extract from the Web (KnowIt, and by Hovy), community knowledge entry (OpenMind), dictionary mining (MindNet, Extended WordNet), statistics over corpora (Schubert), acquire from users with nice knowledge acquisition tools (RKF, Vulcan).]

• On WordNet, Text Mining, and Knowlege Bases of the Future (Peter Clark, 2005, invited talk to Oren Etzioni's students at Univ Washington)

  Abstract:
  In this informal talk, I'll discuss two closely related topics:
  

  • At Boeing, we have been attempting to use WordNet for machine reasoning, using both its taxonomic information and other information (e.g., parts relations). While we get some leverage, it is clear that WordNet is drastically limited in the types of knowledge it contains. I will describe our work in this area, and present a vision for what we believe a future WordNet-like resource should look like -- a large knowledge base with vastly richer connectivity between concepts -- and the amazing potential such a resource would offer for machine reasoning, if someone were to create it.
    
  • Back in 2002, Len Schubert made a conjecture that: beneath the explicit surface level of text there lies a largely untapped source of general knowledge. He also proposed and demonstrated methods for mining text to tap into that general world knowledge. At Boeing we have reimplemented (a version) of his approach, and extracted 22 million simple common-sense "facts" from 1GB of newswire text. I'll describe this work, and what we learned from it.