<?xml version="1.0" standalone="yes"?>
<Paper uid="H93-1007">
  <Title>THE SEMANTIC LINKER- A NEW FRAGMENT COMBINING METHOD</Title>
  <Section position="8" start_page="41" end_page="41" type="concl">
    <SectionTitle>
6. RESULTS AND DISCUSSION
</SectionTitle>
    <Paragraph position="0"> Results from the November 1992 DARPA evaluation[6] show that the Semantic Linker reduced DELPHI's Weighted Error rate on the NL-only portion of the test by 30% (from 32% to 22%). This was achieved mostly by dramaticaly lowering the No Answer rate (from 21% to 8%).</Paragraph>
    <Paragraph position="1"> It should be noted that these results were achieved with an earlier w~rsion of the Semantic Linker than that reported here. In particular, this earlier version did not make use of empirically determined probabilities, but rather used a more ad hoe system of heuristically determined weights and features. Nevertheless, these preliminary results give us some confidence in our approach.</Paragraph>
    <Paragraph position="2"> Several areas of future work are seen. One is the use of automatic training methods to determine feature weights. A corpus pairing sentences and sets of connecting links could be used in supervised training to adjust initial values of these weights up or down.</Paragraph>
    <Paragraph position="3"> Another area, one in which we are already engaged, is using the Semantic Linker in ellipsis processing by treating the preceding utterance as a fragment-structure into which to link the present, elliptical one.</Paragraph>
    <Paragraph position="4"> A third area of future work is the use of relational probabilities and search in the generation of fragments themselves. Currently, the fragment generator component is entirely separate from the rest of the Linker, which makes it diflicdt for combination search to recover from fragment generation. Instead of trying to combine fragments, the Linker could seek to combine the semantic objects internal to them, in a process where inter-object links found by the fragment generator would have a strong but not insurmountable advantages A last area of future work is to more fully integrate the Semantic Linker into the regular parsing mechanism itself, and to investigate ways in which parsing can be viewed as similar to the linking process.</Paragraph>
  </Section>
class="xml-element"></Paper>