<?xml version="1.0" standalone="yes"?>
<Paper uid="W93-0303">
  <Title>Document Filtering Using Semantic Information from s Machine Readable Dictionary1</Title>
  <Section position="6" start_page="26" end_page="26" type="evalu">
    <SectionTitle>
6. Testino and Results
</SectionTitle>
    <Paragraph position="0"> Having produced a ranked listing of documents based on the similarity of their SFC vectors to a query vector, the most illustrative evaluation of performance would be the results provided in Table 1. We believe that these are quite reasonable filtering results. Earlier testings of the SFCoder have revealed that the most important factor in improving its performance would be recognition that a query contains a requirement that a particular proper noun be in a document in order for the document to be relevant.</Paragraph>
    <Paragraph position="1"> Therefore, we have incorporated a second level of lexical-semantic processing as an extension of the SFCoder. That is, the Proper Noun Interpreter (Peik et el; in press) includes algorithms for computing the similarity between a query's proper noun requirements and each document's Proper Noun Field. The proper noun similarity value is then combined with the similarity value produced by the SFCoder for a re-ranking in relation to the query. In the 18th month TIPSTER evaluation of our system, this re-ranking of documents based on the SFC values plus the Proper Noun values improved significantly the filtering power of the system. We have not yet adapted the PSV for predicting the cut-off criterion on the combined similarity values, but we will be doing so in the next few weeks.</Paragraph>
  </Section>
class="xml-element"></Paper>