<?xml version="1.0" standalone="yes"?>
<Paper uid="W94-0114">
  <Title>Statistical versus symbolic parsing for captioned-information retrieval</Title>
  <Section position="2" start_page="0" end_page="112" type="intro">
    <SectionTitle>
2. Introduction
</SectionTitle>
    <Paragraph position="0"> Our MARIE project has been investigating information retrieval of multimedia data using a new idea: putting primary emphasis on caption processing.</Paragraph>
    <Paragraph position="1"> Although content analysis methods such as substring searching for text media and shape matching for picture media can obviate captions, content analysis usually requires unacceptably-large amounts of time at retrieval time. Captions can be cachings of the results of content analysis, but they can also include auxiliary information like the date or customer for a photograph.</Paragraph>
    <Paragraph position="2"> Since captions can be considerably smaller than the media-data they describe, checking captions before retrieving media data can save time if it can rule out many bad matches quickly, the idea of &amp;quot;information filters&amp;quot; (Belkin and Croft, 1992).</Paragraph>
    <Paragraph position="3"> However, caption processing does not necessarily give faster multimedia retrieval. The terms of the caption are often synonyms or subterms of those supplied by a user during retrieval, so a complete thesaurus of synonyms and a complete type hierarchy of terms should be used during information retrieval (Smith et al, 1989). Furthermore, to obtain high query recall and precision, natural-language processing of the captions must be done to determine the word senses and how the words relate, to get beyond the well-known limits of   -2keyword matching (Krovetz and Croft, 1992). This additional processing could be slow, so the MARIE project is concerned with methods of improving its efficiency in caption-based retrieval.</Paragraph>
    <Paragraph position="4"> This paper reports on an important direction that we have explored recently: mixing traditional symbolic parsing with probabilistic ranking based on a restricted kind of statistical information. While the MARIE project is intended for multimedia information retrieval in general, we have used as testbed the Photo Lab of the Naval Air Warfare Center (NAWC-WD), China Lake California USA. This is a library of approximately 100,000 pictures and 37,000 captions for those pictures. The pictures cover all activities of the center, including pictures of equipment, tests of equipment, administrative documentation, site visits, and public relations. With so many pictures, many of which looking virtually identical, captions are indispensable to find anything. But the existing computerized keyword system for finding pictures from their captions is unhelpful, and is mostly ignoredby personnel. (Rowe and Guglielmo, 1993) reports on MARIE-l, a prototype implementation in Prolog that we developed for them, a system that appears much more in the direction of what users want.</Paragraph>
    <Paragraph position="5"> But MARIE-1 took a man-year to construct and only handled 220 pictures (averaging 20 words per caption) from the database. To handle the full database, efficiency and implementationdifficulty concerns have become paramount. MARIE-2, currently under development, will address these problems by exploiting a large statisticalcorrelation database, allowing for simpler parse rules and fewer semantic routines. This should make it run more efficiently while being much easier to apply to the full captions database. This will provide an interesting test of statistical parsing ideas from an engineering standpoint.</Paragraph>
  </Section>
class="xml-element"></Paper>