A corpus-based knowledge representation system consists of a large collection of disparate knowledge fragments or schemas, and a rich set of statistics computed over the corpus. We argue that by collecting such a corpus and computing the appropriate statistics, corpus-based representation offers an alternative to traditional knowledge representation for a broad class of applications. The key advantage of corpus based representation is that we avoid the laborious process of building a (often brittle) knowledge base. We describe the basic building blocks of a corpus-based representation system and a set of applications for which such a paradigm is appropriate, including one application where the approach is already showing promising results.
It's a six page paper that was an invited talk at the conference.
The paper is available to members of the Seminar group; see your email or the list archives.