Pattern -oriented access to document collections

This dissertation investigates pattern-oriented access to collections of unstructured text documents. A pattern-oriented information search differs from a more traditional record-oriented search just as the study of an entire forest differs from the inspection of specific trees. For example, to enjoy Abraham Lincoln\u27s eloquence, we might look up a particular speech such as the Gettysburg Address (a trees-perspective); to understand the evolution of Lincoln\u27s ideas, we must seek trends across the collection of his public statements (a forest perspective). Data-mining seeks this forest-perspective by finding statistical patterns in data. Unfortunately, data-mining is only applied to highly-structured data, and therefore ignores much, if not most, of the world\u27s information, which exists as unstructured text. Evidence from the Information Retrieval, Information Visualization, Bibliometrics, and Library Science literatures demonstrate that pattern-oriented access to document collections is a critically important task; one in which people often engage even if they do not have tools designed for this purpose. Informed by these literatures, a prototypical pattern-discovery system named Homer is introduced and applied in two empirical studies. The first study required subjects to answer specific questions about the prose of a photographer\u27s captions; the second study required subjects to respond to open-ended medical questions based on a collection of emergency room medical reports. Results show Homer users learning more and taking less time, on average, than users of more-traditional record-oriented systems. These results, combined with evidence from the literature, argue strongly that pattern-oriented access to document collections is possible, and can potentially tap vast, previously-unavailable sources of knowledge by helping us find the stories hidden within our document collections